Construction of the axillary lymph node atlas of BC
Using 3D-reconstructed images from 43 patients, we constructed a new axillary lymph node atlas. In some patients who underwent PET-CT, we compared the PET-CT and 3D reconstruction images from two patients: one patient had axillary lymph node metastasis, and the other had no metastasis (Fig. 2). In patients without lymph node metastasis, the shape of the lymph nodes was regular, and most of the lymph nodes were ellipsoidal (Fig. 2A-B). In patients with lymph node metastasis, the shape of the lymph nodes began to become irregular, and most lymph nodes changed from oval to more spherical (Fig. 2C-D).
Formulas and models for predicting lymph node metastasis in the training cohort
After determining that the shape of the metastatic lymph nodes may be more spherical than oval, we suspected that the difference in sphericity between the lymph nodes may help us to identify diseased lymph nodes. In a training cohort containing 16 patients, we measured the sphericity of a total of 137 lymph nodes, 37 of which had metastases and 100 of which did not. Using the variation in sphericity of the lymph nodes, we created a sphericity formula. The ROC curve for this formula was plotted by substituting the a, b and c values into the sphericity formula, with an AUC of 0.773 (Fig. 3D), a sensitivity of 91.9% and a specificity of 52.0% (Fig. 4A-B, Table S1), with significantly higher values in the metastatic group than in the nonmetastatic group (Fig. 3C). We also used a 2D plane-based formula, which uses the ratio of the short diameter to the long diameter to identify metastatic lymph nodes, with an AUC of 0.604 (Fig. 3B), a sensitivity of 75.5% and a specificity of 46.3% (Fig. 4A-B, Table S1), all of which were lower than those obtained from the sphericity formula, with no significant difference between the metastatic and nonmetastatic groups (Fig. 3A). Thus, the diagnostic performance of the 3D formula was significantly better than that of the 2D formula.
To further improve diagnostic efficacy, we attempted predictive classification using a decision tree model. As shown in Figure S2, the 137 lymph node samples from the training set were divided into nodes according to the values of the independent variables a, b and c so that the distribution of the dependent variable within the terminal nodes converged as closely as possible. The model had a total of 19 prediction errors, of which 16 were misdiagnosed and three were missed, with an accuracy of 86.1%, a sensitivity of 68.0% and a specificity of 96.6% (Table S1). We also ranked the importance of the three predictor variables, with the target variable c being the most important, followed by a (Figure S3A).
On the basis of the decision tree model, we used an algorithm called random forest to establish a random forest model to further improve prediction efficiency. At present, the random forest algorithm has superior classification performance. The random forest prediction model had an AUC value of 0.844, a sensitivity of 88.9% and a specificity of 80.0% (Figure S3B). It had the lowest error rate when there were 100 decision trees, and the prediction variable a became more important in this model (Figure S3C-D). Among the 137 samples, there were only 5 prediction errors, the accuracy was as high as 96.3%, and the performance was greatly improved compared with that of the decision tree model.
Overall validation in the validation cohort
This study provides an overall validation of the methods of each study. In the validation cohort, the sphericity formula had a much higher correct classification rate of 86% compared that of the 2D formula (69.8%). In the decision tree model and random forest model, the correct classification rate improved to 88.4% and 90.7%, respectively, which were significantly different compared to those of the 2D formula (p = 0.031; p = 0.014, Table S3).
Next, to further assess the performance of each method, we compared each diagnostic method with ultrasound. A total of 39 patients underwent preoperative 3D reconstruction and ultrasound. Ultrasound showed no metastases in 7 patients, lymph node enlargement in 15 patients, lymph node supraphysiology in 6 patients, metastatic lymph nodes in 10 patients and heterogeneous lymph nodes in 1 patient. Compared with the pathological findings, ultrasound correctly diagnosed 7 cases of metastatic lymph nodes and 8 cases of nonmetastatic lymph nodes; ultrasound missed the diagnosis in 21 patients and misdiagnosed 3 patients, with a correct classification rate of only 38.5%, which was significantly lower than that of each diagnostic method (Fig. 4C, Tables S2-3).
Finally, we attempted an evaluation with CT. The correct classification rate was 48.8%, and this rate was significantly different from those of the other diagnostic methods (Fig. 4C, Tables S2-3).
Combined diagnostic method is better than a single diagnostic method
After assessing the diagnostic efficacy of each method, we speculated whether the combination of two methods would further improve diagnostic accuracy; here, we combined the sphericity formula with ultrasound and CT. We found that the sphericity formula had a sensitivity of 96.4% and a specificity of 81.8% when combined with ultrasound for diagnosis. When combined with CT, it had a sensitivity of 93.5% and a specificity of 83.3%, with high diagnostic efficacy (Fig. 4A-B). Upon validation in 39 patients, ultrasound missed the diagnosis in one patient and misdiagnosed two patients, with a correct classification rate of 92.3%; in the total validation set, ultrasound missed the diagnosis in 2 patients and misdiagnosed 2 patients, with a correct classification rate of 90.7% when the sphericity formula was combined with CT (Fig. 4C, Table S2). This result suggests that diagnostic efficacy is improved when diagnostic methods are combined.
Assessment of lymph nodes after NAC
Considering that the assessment of axillary lymph nodes after NAC is also an important part of subsequent treatment decisions, we assessed lymph nodes in patients after NAC using the methods described above. We included a total of 11 patients, five of whom had lymph node metastases and six of whom did not, as determined by the pathological findings post neoadjuvant surgery. These patients underwent 3D CT reconstruction before and after NAC. The morphology of the lymph nodes changed significantly before and after NAC (Fig. 5A-B). We attempted to measure the sphericity of the patients' lymph nodes after NAC in the 3D reconstruction system, and a total of 90 lymph nodes were measured. Among the various assessment methods, ultrasound had a low correct classification rate of 45.5%, followed by the 2D formula and CT (54.5%); the sphericity formula had a better correct classification rate than the 2D method (63.6%); and the decision tree and random forest models and the combined diagnostic methods had the highest correct classification rate (72.7%) (Fig. 5C, Table S4-5). Overall, the diagnostic efficacy of the 3D method and the combined diagnostic method was better than that of the 2D method, but the diagnostic efficacy decreased in patients not receiving NAC.