Using Artificial Intelligence for Analysis of Histological and Morphological Diversity in Salivary Gland Tumors

doi:10.21203/rs.3.rs-1966782/v1

Download PDF

Article

Using Artificial Intelligence for Analysis of Histological and Morphological Diversity in Salivary Gland Tumors

https://doi.org/10.21203/rs.3.rs-1966782/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Salivary gland tumors (SGT) are aheterogeneous neoplasms with large morphological diversity and overlapping features. Recently, numerous artificial intelligence (AI) methods shown for reproducible histological diagnosis and prognosis. However, their application to SGT has not been reported to date.

This study aims to examine if AI can be used to differentiate between different SGT subtypes based on the analysis of digitized whole-slide images (WSIs) of Haematoxylin and Eosin (H&E) stained slides.

A two-stage machine learning (ML) algorithm was developed and tested on 240 scanned H&E WSIs of SGT cases using an open-source bioimage analysis software (QuPath) to train and analyze features on representative regions of interest. The first classifier was designed to differentiate between two benign and four malignant SGT subtypes with an equal split between benign and malignant SGTs (n = 120 each), while the second classifier was used for malignant SGT subtyping (n = 120). Features extracted using the ML classifiers were also analysed using deep learning (DL) networks to determine any performance improvements.

Our first classifier showed excellent accuracy for automated differentiation between benign and malignant SGTs (F1-score = 0.90). The second classifier also performed well for differentiation between four different malignant SGTs (average F1 = 0.92). Significant differences between cellularity, nuclear hematoxylin, cytoplasmic eosin, and nucleus/cell ratio (p < 0.05) were seen between tumors in both experiments. Most of the DL networks also achieved high F1-scores for benign versus malignant differentiation (> 0.80), with EfficientNet-B0 giving the best performance (F1 = 0.87) but with inferior accuracy than the ML classifier for malignant subtyping (highest F1 = 0.60 for ResNet-18 and ResNet-50).

Our novel findings show that AI can be used for automated differentiation between benign and malignant SGT and tumor subtyping on H&E images. Analysis of a larger multicentre cohort using ML and DL at the WSI level is required to establish the significance and clinical usefulness of these findings.

Salivary gland tumors

Artificial Intelligence

Machine Learning

Deep Learning

salivary gland cancer

QuPath

image analysis

histopathology

Salivary gland tumors (SGTs) are a heterogeneous group of neoplasms that present with a wide range of histological features and subtypes. They are rare, with an annual incidence of approximately 2.5–3.0 per 100,000 people in the Western world (1). SGTs can be diagnostically challenging due to a large number of entities and markedly similar and overlapping features but different clinical behavior (2). A large number of additional stains, immunohistochemistry (IHC), and molecular work (Fluorescence in Situ Hybridization (FISH), Polymerase Chain Reaction (PCR), etc.) can be required to diagnose and sub-type SGT. However, these tests are available only at specialist centers with associated time delays and cost implications. In some cases, even when these investigations can be performed, differentiation between tumor types can be challenging (3), and unfortunately, very few of these special tests are available in the developing world.

Furthermore, contextual and morphological information is not evident when molecular or genetic tests are performed, and a number of SGTs have no known molecular alterations at present. Identification of the tumor type is important as it provides guidance for the optimal management and prognostic behavior. These issues highlight the need for novel, efficient, and widely available methods as diagnostic aids.

The advent of digital pathology and the ability to obtain whole slide images (WSIs) from histology slides has accelerated the application of artificial intelligence (AI) to pathology, which has been helped by advancements in computing power and technology, allowing exploration of sub-visual morphometric features with a potential to improve patient care (4). Numerous reports have shown the use of AI for consistent and quantitative histological diagnosis as well as providing prognostic information in a range of cancers (5–10). However, its application to SGTs has not been investigated or reported to date.

Machine learning (ML) is a branch of AI that learns from data that has been provided to make a prediction (11–13). ML techniques have been employed by several image analysis platforms to analyze cell, tissue and histology images. Open-source image analysis platforms with built-in ML capabilities (such as QuPath) have been utilized for histological feature analysis and staining quantification (14).

Artificial neural networks (ANNs) consist of numerous fully connected layers, including an input layer, an output layer, and various hidden layers in between. Neural networks, therefore, model the way the brain's neural network system works to recognize patterns in data (13). Deep learning (DL) encompasses a range of approaches that are based on ANNs, but typically with many hidden layers. Convolutional neural networks (CNNs) are a DL approach that have become dominant in the computer vision and computational pathology literature for both segmentation and classification tasks, owing to their efficiency on inference and generalisability to new data (4).

To the best of our knowledge, ML and DL tools have not been used for SGT histological analysis to date. In this study, we first aim to use ML to differentiate between benign and malignant SGTs, and to subtype malignant SGTs in WSIs (both of which can be challenging tasks in diagnostic practice). For this purpose, QuPath was used for both feature generation and classification within manually selected regions of interest (ROIs). After the initial analysis within QuPath SGT classification was also performed using established Python libraries outside QuPath. Finally, owing to the recent popularity and success of DL techniques, these results were compared with DL networks applied to the same tasks.

Benign and malignant SGTs were identified using a digital database and the corresponding Hematoxylin & Eosin (H&E) stained slides were retrieved from the department archive (Unit of Oral and Maxillofacial Pathology, School of Clinical Dentistry, University of Sheffield, UK- Health and Care Research Wales ethics reference: 20/WS/0017). This department receives a large number of regional, national and international consult cases for expert opinion, and both internally and externally stained slides were included in the cohort to ensure robustness of training. The retrieved cases included two common benign SGT subtypes, i.e., pleomorphic adenoma (PA) and basal cell adenoma (BCA), and four malignant SGT subtypes, i.e., mucoepidermoid carcinoma (MEC), adenoid cystic carcinoma (AdCC), acinic cell carcinoma (ACC) and carcinoma ex pleomorphic adenoma (Ca-ex-PA). The diagnosis of these cases was confirmed, followed by anonymization of the slides. Whole slide images (WSIs) were generated using an Aperio CS2 scanner (Leica Biosystems, Nussloch, Germany) at 40x magnification. Calibration was done prior to each scanning session and images were stored on a dedicated server. Anonymised WSIs of cases to be analyzed were downloaded from the server for analysis.

An open-source bioimage analysis software (QuPath) was employed (QuPath is freely available to download for Windows, Mac, and Linux operating systems and supports most of the known WSI and image formats through Bio-Formats and OpenSlide libraries). In addition to cell detection and IHC quantification, it also allows users (including pathologists) to perform training on their own datasets and build ML algorithms for automated identification (classification) into different cell/tissue subtypes (14) as well as the location and delineation of these features (segmentation). A wide range of features were analyzed and compared between tumors, including cellularity, nuclear and cytoplasmic features such as circularity, eccentricity, hematoxylin and eosin (H&E) optical density and nucleus/cell area ratio.

Case Selection and Building of Machine Learning Classifiers

For comparison between benign and malignant SGT, WSIs of H&E stained sections were used, including 120 from benign tumors (PA and BCA) and 120 from malignant tumors (including MEC, AdCC, ACC, and Ca-ex-PA). We used 67% (n=160) of the cases for training and the remaining for testing (n=80). The training was performed on an equal number of benign and malignant cases (i.e., n=80 WSI for each category, 160 in total). To train the benign vs malignant (BvM) detection classifier, ROIs per WSI were selected using fixed-size areas of 142,884 μm²(1,500 × 1,500 pixels) to ensure standardization across cases. Next, cell detection analysis was performed, following which the detected cells/nuclei were assigned to a specific class/ground truth (i.e., benign or malignant based on the histological review/diagnosis or the appropriate subtype). At least five different ROIs were used for training from each WSI, ensuring that morphologically different tumor-rich areas were included, e.g., cribriform pattern, clear cell areas, as well as solid or tubular patterns (where applicable). Using the ROIs in the training cases, a ML classifier (Random Forest/RF) was built and validated through visualization of nuclear segmentation in the unseen cases. 80 unseen WSIs with 400 ROIs of fixed-size areas were used to blindly test and validate the BvM classifier for automated classification, followed by quantification and statistical analysis of color and morphometric features leading to the classifier’s 'decision' (Figure 1) (Supplementary Table 1).

The second part of the study aimed to build an additional classifier for malignant tumor subtyping (MST) for automated identification of the more common malignant SGT. For this part of the study, 120 WSIs were used for training and testing. Although these cases were the same as those used in the previous part of the study, a new classifier was built to ensure that the test cases remained 'unseen'. Two-thirds (n=80) WSIs were used for training and the remainder for testing (n=40). A RF classifier was trained using 80 WSIs with 400 ROIs, including four different tumors (MEC, AdCC, ACC, and Ca-ex-PA) (n=20 WSIs of each tumor). The ROI dimensions were maintained at 1,500 x 1,500 pixels, and cell/nuclear detection was performed as described previously. All detected cells in each tumor type ROI were assigned to a specific tumor type class (i.e., MEC, AdCC, ACC, Ca-ex-PA). 40 unseen WSIs (10 of each SGT) with 200 ROIs were used to test the MST classifier and to perform analysis and quantification of features (Figure 1).

Comparison with Customized Machine Learning Models

We additionally aimed to understand whether using optimized ML methods (outside of the QuPath environment) would provide superior performance across the two tasks (i.e., malignant vs benign, and tumor subtyping). Since all of the previous work was performed in QuPath (e.g., nuclei segmentation, feature extraction, classification), the performance of the QuPath based RF classifier was compared with the same classifier trained/tested outside QuPath using the Scikit-learn 1.0.1 toolbox with Python (15). The above two tasks were repeated by extracting the QuPath-generated nuclear segmentation and features followed by optimization of the RF models on the training cohort before testing them on the unseen test set.

Comparison with Deep Learning Networks

Next, the utility of DL methods for classification was analyzed. For direct prediction with DL, the ROIs were tessellated into smaller patches (256 x 256, at 40X) before training/testing multiple state-of-the-art convolutional neural networks (CNNs) for automated prediction. Here, we used ResNet-18, ResNet-50, Efficient-NetB0 and Efficient-NetB3 models, built with PyTorch 1.10 (16, 17). On inference, the maximum argument was taken across all patches per subject to achieve predictions.

In this work, we have therefore performed three different sets of experiments:

Feature generation using QuPath and employing the built in RF for classification
Feature generation using QuPath followed by using an optimized RF (outside the platform) for classification
DL for classification (based on raw image patches)

Spatial analysis

Dynamic interactions between tissue features result in topographical characteristics that can explain relationships between different structures. The spatial analysis was performed using a set of features related to the orientation of objects at a certain location. This included proximity by measuring centroid distance between cells as well as cluster Delaunay analysis, including neighbouring cells, intercellular distance, and mean triangle area. Delaunay triangulation is a geometric calculation indicating a set of points that can be found in optimal time and position (18).

Statistical Analysis

T-Test and multiple comparison one-way ANOVA were used to measure statistical significance between different geometrical, spatial, and staining features. Microsoft Excel 2016 (Microsoft Office Software, USA) was used to organize exported data and perform statistical analyses. The performance of detection classifiers was measured using precision, recall, F1 score, and AUROC, generated at the case level.

Precision = True positive / true positive + false positive

Recall = True positive / true positive + false negative

F1 score = 2x (precision x recall/ precision + recall)

AUROC= (x-axis) 1 – specificity (= false positive fraction = FP/(FP+TN)), (y-axis) sensitivity (= true positive fraction = TP/(TP+FN))

1. Performance of the classifiers

The benign versus malignant tumor RF classifier (BVM) showed highF1 and AUROC scores of 0.90 (Table 1). The predominant detection/class was accurate for all cases, although some false positive detections for both benign and malignant cells were identified (Figure 2). The customized RF classifier built outside QuPath (using different tree depth and parameters) gave a superior performance (F1 = 0.95, AUROC - 0.95, Table 1). Results from DL were similar, with all CNNs showing high F1-scores (>0.80, Table 1), with EfficientNet-B0 giving the best performance (F1= 0.87). However, none of the DL CNNs surpassed the performance achieved using the customized RF ML classifier (F1=0.95) (Table 1).

The malignant subtyping QuPath classifier (MST) also showed excellent performance for classification and automatic differentiation between SGT subtypes (including MEC, AdCC, ACC and Ca-ex-PA), with the predominant automated detection correct in all instances (F1 = 0.92, Table 1). However, there were some false-positive detections as shown in Figure 2. Similar to the previous task, the customized RF classifier built outside QuPath (but using QuPath generated features), achieved the best performance (F1 = 0.95, Table 1). Interestingly, results for DL networks appeared inferior to both ML classifiers (highest F1 score=0.60 with ResNet-18 and ResNet-50).

2. Features contributing to the classifier performance

The geometrical and morphometrical feature analysis of the test set demonstrated a number of features that were being used by the BVM classifier for benign or malignant class prediction. Statistical analysis showed that there was a significant difference between nuclear circularity, nuclear hematoxylin optical density (OD), and cytoplasmic eosin OD between benign and malignant tumors (p<0.01). Furthermore, a significant difference in nucleus/cell ratio was also seen (p<0.05). No difference was seen between benign and malignant SGT for nuclear eccentricity (Figure 3). Similar to results seen for the BVM classifier, a detailed analysis of features (listed in Supplementary Table 1) showed various geometrical and morphometrical features guiding the automated detection of malignant subtypes. There was a significant difference between nuclear hematoxylin OD, cytoplasmic and eosin OD, and nucleus/cell ratio (p<0.01) between the different malignant tumors. Interestingly, no statistically significant difference was seen between nuclear circularity and eccentricity for this task (Figure 4).

3. Quantitative morphometrical feature analysis

Cellularity

The average number of cells in malignant tumors (mean= 6490, SD= ±1720) was higher than benign tumors (Mean=5087, SD= ±2659) (average cellularity per case calculated across five standardized tumour rich ROIs). This difference in cellularity was statistically significant (p<0.01) (Figure 3B). Comparison of malignant tumours showed the highest cellularity in Ca ex PA (Mean=7428, SD=±1474 cells), followed by AdCC (Mean=6454, SD=±1863) and ACC (Mean=6187, SD= ±1395). MEC demonstrated lower cellularity compared to other SGT (Mean=4875, SD=±1040 cells). Statistical analysis showed that the difference in cellularity between the four SGT was statistically significant (p<0.01) (Figure 4).

Spatial analysis

The average centroid distance showed a significant difference between cells in malignant and benign SGT, with a smaller distance seen in malignant SGT (p<0.01) (Figure 5C). Cluster spatial analysis showed a significant difference in Delaunay neighbouring cells in benign and malignant SGT (p<0.01). Furthermore, there was a significant difference between the Delaunay mean intercellular distance between benign (Mean=14.39μm, SD=±2.76) and malignant (Mean=12.20μm, SD=±1.70) tumors (p<0.01). In addition, the Delaunay mean triangle area was also significantly different between benign (Mean=73.25, SD=±13.59) and malignant (Mean=56.58, SD=±15.65) SGT (p<0.01) (Figure 5D).

Spatial analysis for different malignant subtypes showed the highest number of Delaunay neighbouring cells in ACC (Mean=4.48, SD= 1.11 cells) compared to MEC (Mean=3.75, SD=±0.85), AdCC (Mean=3.52, SD=±0.87) and Ca-ex-PA (Mean=3.31, SD=±0.78). This difference was statistically significant between the different SGT (p<0.05). The Delaunay mean distance was also significantly different between different types of malignant SGT (p<0.01). In addition, the Delaunay mean triangle area also showed a significant difference (p<0.01) between subtypes with the highest area seen in MEC (Mean=72.62, SD=±12.37) and the lowest in AdCC (Mean=54.30, SD=±15.98) (Figure 6 B).

Computational pathology (ML, DL, and AI) has been shown to provide objective and accurate quantification and classification in a range of cancers (5–9, 19). Open-source image analysis software (such as QuPath) have built-in ML tools which can be used to train and generate a few types of classifiers/algorithms that have the ability to perform automated cell detection, classification, segmentation, and automated measurements of objects after appropriate training (20).

Our novel results report the application of AI and ML for histological analysis and subtyping of SGT on H&E images for the first time. Our ML benign vs malignant (BVM) classifiers (QuPath RF and custom RF) have achieved excellent accuracy with F1 and AUROC values of 0.90 and 0.95, respectively. There is a scarcity of studies that use ML or DL to differentiate between benign and malignant lesions in the head and neck region using histological image analysis (13, 21). Furthermore, to date, the application of these tools for SGT analysis, diagnosis, and sub-typing has not been explored. The only study to date has reported tumor detection in thyroid and salivary tumors on gross macroscopic specimens using hyperspectral images (22). Therefore, a direct comparison of our findings with relevant studies is not possible. Nonetheless, comparable concepts have been explored in a number of studies, For example, one study in breast tissue attempted to distinguish between non-carcinoma (normal and benign) and carcinoma (in situ and invasive) and reached an accuracy of 83.3% (19). Another study of benign and malignant breast lesions using a class structure-based deep learning model (CSDCNN) showed an excellent performance (93.2% accuracy) (23). In prostate adenocarcinoma, an AUC of 0.98 and 0.99, respectively, have been reached using DL to distinguish between benign and malignant tumors on large datasets (24, 25). In addition, excellent results have been reported using CNNs and recurrent neural networks (RNNs) for classification between gastric and colonic adenocarcinoma and adenoma (AUCs of 0.96-97 and 0.99, respectively) (26). Although we have shown novel and interesting data using ML classifiers and DL models for SGT analysis, it is apparent that modern DL techniques can offer significantly higher performance and accuracy at a WSI level on larger datasets, and we are already exploring this as an extension to the existing work. However, SGT are much rarer compared to breast, prostate, and other adenocarcinomas; therefore, curating large multicentre datasets (similar to published studies with large scale validation) is extremely challenging.

Malignant SGT subtyping usually requires the use of several further IHC stains and referral to specialist pathologists with associated time and cost implications as well as variation in availability of these tests, in particular in developing countries. Our novel ML classifier (MST) showed an outstanding performance; however, due to time and data restraints, only four common types of malignant SGT could be included for comparison; these tumors were selected as they account for the majority of malignant SGT in reported studies (27, 28). Our MST classifiers (QuPath RF and custom RF) showed excellent accuracy (F1 scores- 0.92 and 0.95, AUROC- 0.95 and 0.97). Interestingly, but not surprisingly, the performance was inferior for DL for this task. DL algorithms require larger datasets, especially for challenging tasks, and it is likely that the use of ROIs and four different classes/tumors in a relatively small dataset was insufficient to achieve good accuracy. Although never reported in SGT, DL has been used to subtype other glandular carcinomas, such as lung Adenocarcinoma (LUAD) vs squamous cell carcinoma (LUSC) or normal lung tissue in much larger datasets at WSI level (AUC of 0.97) (29).

Although ML offers the advantage of automated analysis of histopathological features, it requires a large amount of training and annotations, making it tedious and time-consuming. DL algorithms can learn directly from raw data but it can be difficult to establish how algorithms make the decision from input data and arrive at a prediction (30). In addition, DL algorithms are data-hungry, often requiring large cohorts. Our study uses a hybrid approach for SGT classification and segmentation, including analysis of morphological, geometrical and colour features that ML uses for decision making and comparing that with DL algorithms (using ML generated ROI level features).

For benign vs malignant SGT, the analysis of the geometrical and morphometrical features demonstrated a significant difference between nuclear circularity, nuclear hematoxylin, cytoplasmic eosin and nucleus/cell ratio. This is biologically relevant as malignant cells are known to have a higher and abnormal DNA content (i.e., hyperchromatism), leading to a higher nuclear hematoxylin staining and nuclear/cell ratio (31). Nuclear hematoxylin staining and nucleus/cell ratio were also statistically different between different malignant SGT, which could be related to the degree of differentiation or histological grade of those tumors.

Assessment of tumor cellularity is a subjective and tedious process, not routinely used for diagnosis or prognosis prediction, and with concerns of inter-observer variability among pathologists. ML and DL can aid this process and have been shown to objectively quantify tumour cellularity (32). Our results show higher cellularity in malignant SGT compared to benign. This observation is in agreement with the literature as malignant tumors are known to be more cellular, but this quantification is rarely done in practice, in particular for SGT (30). Yin et al. (2016) developed a quantification algorithm for tumor cellularity and found that cancer cells show higher cellularity than non-cancers in lung tissues (33). Our cellularity data comprised an analysis of five fixed-size ROIs in tumor-rich areas. It would be interesting to explore tumor and stroma separately and at the WSI level instead of ROIs to further determine the significance of these findings.

Our paper is also the first to report the importance of spatial characteristics in SGT. This is important as the tumor microenvironment has been shown to play a key role in progression and prognosis of numerous neoplasms. Furthermore, in many types of cancers, histopathological analysis can show cells growing in clusters and show architectural patterns and organization. Ali et al., (2013) reported a cell cluster graph (CCG) which computationally characterized prostate cancer tissue images according to spatial distribution and correlated it with histological grading (34). Our study shows that spatial orientation and clustering of cells are different between benign and malignant tumors as well as different malignant subtypes and this might be a valuable adjunct for differentiating between them.

Despite the novelty of our findings, there are a number of shortcomings. Our analysis mainly focused on epithelial cells at the ROI level. We are already exploring stromal features and analysis at the WSI level as an extension to this work. Furthermore, our dataset included cases from a single department. However, this is a center of excellence with regards to SGT analysis, receiving routine and referral cases from not only across the region but also nationally and internationally, which would mitigate bias and add to the robustness of the trained algorithms.

To conclude, AI has an enormous potential to aid diagnosis and improve patient care. Our novel findings show that AI can be used for analysis, quantification, and differentiation between salivary gland tumors. A larger multicentre cohort needs to be analyzed to determine the true significance and clinical usefulness of these findings.

Acknowledgments.

N/A

Ethics Approval.

Health and Care Research Wales ethics reference: 20/WS/0017

Author Contributions.

IAS scanned the slides, performed the training, built/tested the ML classifiers, performed statistical analysis and led the writing of the manuscript. NA contributed to feature extraction, analysis and manuscript preparation, AS performed the custom ML analysis outside QuPath as well as the DL analysis. NMR provided guidance with ML and DL approaches and manuscript preparation. SAK conceived the study, led on the ethical approval, case selection, data curation and management, analysis and writing up.

Funding.

The first author’s PhD has been funded by King Saud University, Riyadh, Saudi-Arabia.

Data Availability Statement

The ML classifiers built and used for the current study are available from the corresponding author on reasonable request.

Speight P, Barrett A. Salivary gland tumours. Oral diseases. 2002;8(5):229–40.
Khurram SA, Barrett AW, Speight PM. Diagnostic difficulties in lesions of the minor salivary glands. Diagnostic Histopathology. 2017;23(6):250–9.
Speight PM, Barrett AW. Salivary gland tumours: diagnostic challenges and an update on the latest WHO classification. Diagnostic Histopathology. 2020;26(4):147–58.
Bera K, Schalper KA, Rimm DL, Velcheti V, Madabhushi A. Artificial intelligence in digital pathology—new tools for diagnosis and precision oncology. Nature reviews Clinical oncology. 2019;16(11):703–15.
Bejnordi BE, Veta M, Van Diest PJ, Van Ginneken B, Karssemeijer N, Litjens G, et al. Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer. Jama. 2017;318(22):2199–210.
Wang S, Zhu Y, Yu L, Chen H, Lin H, Wan X, et al. RMDL: Recalibrated multi-instance deep learning for whole slide gastric image classification. Medical image analysis. 2019;58:101549.
Li W, Li J, Sarma KV, Ho KC, Shen S, Knudsen BS, et al. Path R-CNN for prostate cancer diagnosis and gleason grading of histological images. IEEE transactions on medical imaging. 2018;38(4):945–54.
Kather JN, Pearson AT, Halama N, Jäger D, Krause J, Loosen SH, et al. Deep learning can predict microsatellite instability directly from histology in gastrointestinal cancer. Nature medicine. 2019;25(7):1054–6.
Shaban M, Khurram SA, Fraz MM, Alsubaie N, Masood I, Mushtaq S, et al. A novel digital score for abundance of tumour infiltrating lymphocytes predicts disease free survival in oral squamous cell carcinoma. Scientific reports. 2019;9(1):1–13.
Kather JN, Heij LR, Grabsch HI, Loeffler C, Echle A, Muti HS, et al. Pan-cancer image-based detection of clinically actionable genetic alterations. Nature cancer. 2020;1(8):789–99.
Zarella MD, Bowman D, Aeffner F, Farahani N, Xthona A, Absar SF, et al. A practical guide to whole slide imaging: a white paper from the digital pathology association. Archives of pathology & laboratory medicine. 2019;143(2):222–34.
Hamet P, Tremblay J. Artificial intelligence in medicine. Metabolism. 2017;69:S36-S40.
Mahmood H, Shaban M, Indave B, Santos-Silva A, Rajpoot N, Khurram S. Use of artificial intelligence in diagnosis of head and neck precancerous and cancerous lesions: A systematic review. Oral Oncology. 2020;110:104885.
Bankhead P, Loughrey MB, Fernández JA, Dombrowski Y, McArt DG, Dunne PD, et al. QuPath: Open source software for digital pathology image analysis. Scientific reports. 2017;7(1):1–7.
Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, et al. Scikit-learn: Machine learning in Python. the Journal of machine Learning research. 2011;12:2825–30.
He K, Zhang X, Ren S, Sun J, editors. Deep residual learning for image recognition. Proceedings of the IEEE conference on computer vision and pattern recognition; 2016.
Tan M, Le Q, editors. Efficientnet: Rethinking model scaling for convolutional neural networks. International conference on machine learning; 2019: PMLR.
Preparata FP, Shamos MI. Computational geometry: an introduction: Springer Science & Business Media; 2012.
Araújo T, Aresta G, Castro E, Rouco J, Aguiar P, Eloy C, et al. Classification of breast cancer histology images using convolutional neural networks. PloS one. 2017;12(6):e0177544.
Bankhead P, Fernández JA, McArt DG, Boyle DP, Li G, Loughrey MB, et al. Integrated tumor identification and automated scoring minimizes pathologist involvement and provides new insights to key biomarkers in breast cancer. Laboratory investigation. 2018;98(1):15–26.
Mahmood H, Shaban M, Rajpoot N, Khurram SA. Artificial Intelligence-based methods in head and neck cancer diagnosis: An overview. British journal of cancer. 2021;124(12):1934–40.
Halicek M, Dormer JD, Little JV, Chen AY, Fei B. Tumor detection of the thyroid and salivary glands using hyperspectral imaging and deep learning. Biomedical Optics Express. 2020;11(3):1383–400.
Han Z, Wei B, Zheng Y, Yin Y, Li K, Li S. Breast cancer multi-classification from histopathological images with structured deep learning model. Scientific reports. 2017;7(1):1–10.
Ström P, Kartasalo K, Olsson H, Solorzano L, Delahunt B, Berney DM, et al. Pathologist-level grading of prostate biopsies with artificial intelligence. arXiv preprint arXiv:190701368. 2019.
Bulten W, Pinckaers H, van Boven H, Vink R, de Bel T, van Ginneken B, et al. Automated deep-learning system for Gleason grading of prostate cancer using biopsies: a diagnostic study. The Lancet Oncology. 2020;21(2):233–41.
Iizuka O, Kanavati F, Kato K, Rambeau M, Arihiro K, Tsuneki M. Deep learning models for histopathological classification of gastric and colonic epithelial tumours. Scientific reports. 2020;10(1):1–11.
Jones A, Craig G, Speight P, Franklin C. The range and demographics of salivary gland tumours diagnosed in a UK population. Oral oncology. 2008;44(4):407–17.
Alsanie I, Rajab S, Cottom H, Adegun O, Agarwal R, Jay A, et al. Distribution and frequency of salivary gland tumours: an international multicenter study. Head and Neck Pathology. 2022:1–12.
Coudray N, Ocampo PS, Sakellaropoulos T, Narula N, Snuderl M, Fenyö D, et al. Classification and mutation prediction from non–small cell lung cancer histopathology images using deep learning. Nature medicine. 2018;24(10):1559–67.
Salvi M, Acharya UR, Molinari F, Meiburger KM. The impact of pre-and post-image processing techniques on deep learning frameworks: A comprehensive review for digital pathology image analysis. Computers in Biology and Medicine. 2021;128:104129.
Jang SJ, Gardner JM, Ro JY. Diagnostic approach and prognostic factors of cancers. Advances in anatomic pathology. 2011;18(2):165–72.
Rakhlin A, Tiulpin A, Shvets AA, Kalinin AA, Iglovikov VI, Nikolenko S, editors. Breast tumor cellularity assessment using deep neural networks. Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops; 2019.
Yin Y, Sedlaczek O, Warth A, González-Vallinas M, Breuhahn K, Vignon-Clementel IE, et al., editors. Quantitative estimation of tumor cellularity based on histology data. 2016 IEEE Nuclear Science Symposium, Medical Imaging Conference and Room-Temperature Semiconductor Detector Workshop (NSS/MIC/RTSD); 2016: IEEE.
Ali S, Veltri R, Epstein JA, Christudass C, Madabhushi A, editors. Cell cluster graph for prediction of biochemical recurrence in prostate cancer patients from tissue microarrays. Medical Imaging 2013: Digital Pathology; 2013: SPIE.

Table 1. Performances/accuracy metrics of the different classifiers at case-level predictions.
Classifier	Precision	Recall	F1-score	AUROC
*Benign vs Malignant (BVM)*
QuPath RF	1.00	0.80	0.90	0.90
Customized RF	0.93	0.98	0.95	0.95
ResNet-18	0.83	0.80	0.81	0.81
ResNet-50	0.85	0.87	0.86	0.86
EfficientNet-B0	0.93	0.82	0.87	0.87
EfficientNet-B3	0.88	0.76	0.81	0.81
*Malignant subtyping (MST)*
QuPath RF	0.94	0.93	0.92	0.95
Customized RF	0.95	0.95	0.95	0.97
ResNet-18	0.60	0.60	0.60	0.73
ResNet-50	0.63	0.61	0.60	0.74
EfficientNet-B0	0.58	0.60	0.56	0.73
EfficientNet-B3	0.55	0.60	0.54	0.73
Note. RF = Random Forest. Both the QuPath (in built) RF classifier and the customized RF classifiers use nuclear features generated with QuPath; the latter was produced outside of QuPath using the Scikit-learn Python toolbox.

There is NO conflict of interest to disclose.

SupplementaryTable1.docx
Supplementary Table 1: Geometrical and morphometric features which can be analysed within QuPath.

Download PDF

Version 1

posted

You are reading this latest preprint version

Using Artificial Intelligence for Analysis of Histological and Morphological Diversity in Salivary Gland Tumors

Status:

Version 1

Abstract

Figures

Introduction

Material And Methods

Results

Discussion

Declarations

References

Tables

Additional Declarations

Supplementary Files

Status:

Version 1