To our best knowledge, a deep learning model has not been used so far to obtain disconnectome maps. The deep learning 3D U-Net model we trained can generate accurate maps of brain dysconnectivity – deep-disconnectomes – derived from a simple binary lesion mask. These generated disconnectivity maps have a clinical predictive power toward long-term impairment following a stroke that is even higher than conventionally computed disconnectome maps. The validation of the deep-disconnectomes’ capacity to accurately predict neuropsychological scores demonstrate its potential to be released as an AI-driven tool for clinical applications.
Although our deep learning model generates disconnectome-like images, it does not capture exactly the same patterns as the conventionally computed disconnectomes. It can be speculated that the 3D U-Net model is potentially getting rid of noise while producing deep-disconnectomes, allowing for more concise information to be forwarded into the UMAP framework, consequently yielding more accurate predictions. Investigating this question, we conducted a paired t-test comparing conventional disconnectomes with their corresponding (derived from the same binary lesion mask) deep-disconnectomes. The results confirmed our speculation, revealing that deep-disconnectomes emphasize the core white-matter structures while exhibiting comparatively lower capacity in capturing short, more detailed, U-shaped fibers (Fig. 2a). Furthermore, we compared the average Euclidean distances of deep-disconnectomes and conventional disconnectomes to all the other points in their respective morphospace (Fig. 3b). We detected that deep-disconnectomes are better segregated within clusters of near data points. This means that deep-disconnectomes with small Euclidean distances (showing similar disconnectivity patterns) are dissociated better than conventional (BCBTookit) disconnectomes. On the contrary, deep-disconnectomes with large Euclidean distances (showing completely different disconnectivity patterns) are less different than conventional disconnectomes (which might not impact the predictive performances as these disconnectivity patterns are already well segregated). The difference between the deep-disconnectome and the conventional disconnectome cannot be driven by the ‘average template disconnectome’ used as a second input to the 3D U-Net model. Since this template is a fixed parameter of the model, the sole variability of deep-disconnectomes originates from individual lesion masks. Generating a deep-disconnectome without the ‘average template disconnectome’ did not produce sufficient results since the amount of information from the binary lesion mask was not enough to train a deep learning model, applying the parameters we used. This is a common challenge that related studies encountered in the past. Augmenting the input information in a 3D U-Net model with additional data sources or other contextual information has been shown to improve the outcomes in various medical imaging applications. Myronenko et al. (2019) demonstrated that the fusion of multiple modalities (T1, T2, FLAIR images) helped capture diverse tumor characteristics, improving the accuracy of the segmentation results of the model. Another example of a multi-model fusion approach was shown by Guasch and colleagues (2020), who developed a method inspired by geophysics that generates accurate images of the brain applying full-wave inversion. This, however, required their full-wave inversion model to contain an a priori acquired model of the skull, without which it would not have converged to a legitimate solution. While in this exemplary case, an individual skull model needs to be provided in order to achieve satisfactory results, in the present case of obtaining deep-disconnectomes, we only need one single ‘average template disconnectome’, which can be used as a constant for the model input. Providing this information alongside the lesion mask through two channels in the 3D U-Net model allows us to obtain individual predictions of the brain disconnectivity much faster than computing such maps using conventional methods, e.g., whole brain tractography.
While producing deep-disconnectomes with a trained AI model makes obtaining the otherwise processing-heavy data faster, the potential of its application goes beyond the advantage of gaining time. They outperform the conventional disconnectomes when applied in a recent framework for neuropsychological score predictions. Even though the obtained deep-disconnectomes overlapped with the conventional disconnectome maps reaching an average accuracy of R²=0.82 (Fig. 2a), the captured information provides valuable insights into a patient’s disconnection pattern and could be used for further follow-up analyses and predictions. For instance, to evaluate the predicted deep-disconnectomes on a use case, we applied the framework of Talozzi et al. (2023) to predict neuropsychological scores for 86 tests covering different categories of cognitive and behavioral outcomes one year following a stroke. Incidentally, the prediction accuracy of neuropsychological scores using deep-disconnectomes exceeded the original predictions that used conventional disconnectome maps (Fig. 3b). This result confirms that deep-disconnectomes can be used for such analyses. Although the ground truth information is not being reproduced with 100% accuracy, the deep learning 3D U-Net model outputs simplified disconnectomes from a binary lesion mask from which seemingly the UMAP algorithm for dimensionality reduction can extract features in a more precise manner. This added precision led to better regression coefficients and, consequently, to better predictions for clinical test outcomes.
Alongside the impressive performance of the simple 3D U-Net model comes the benefit of time and computational capacity needed to obtain a deep-disconnectome from a lesion mask, making it feasible and practical at the clinical level. This is a common advantage of deep-learning methods, which require a time-demanding training phase with numerous data, but have the tremendous gain of being fast when applied to new instances: a perfect fit for precision medicine suitable for clinical implementation. Using a trained deep-learning model requires the operator to have less than 10 MB of free space on their device, and a deep-disconnectome can be outputted within less than one second. This is an improvement in comparison to the current ways of computing brain disconnections using, e.g., toolkits such as the BCBToolkit (Foulon et al., 2018), the Network Modification (NeMo) Tool Lite (Kuceyeski et al., 2013), or the Lesion Quantification Toolkit (LQT) (Griffis et al., 2021). Usually, programs that are able to compute brain disconnectomes require at least 1,5 GB up to 10 GB of free memory space, taking approximately 720 times longer to create an output (0.642 seconds for the deep-disconnectome in comparison to 480 seconds for the conventional disconnectome). Moreover, the operator needs to follow precise pipelines when reconstructing conventional disconnectomes. The sequence of numerous command steps often leaves room for the erroneous setting of parameters (e.g., the threshold used) or software misuse (e.g., wrong templates for neuroimaging). Not only does the trained 3D U-Net offer a faster and more reliable way of obtaining information about brain disconnectivity, but it also yields more accurate predictions of individual clinical scores.
Finally, a deep learning model that creates disconnectomes based on minimal individual information circumvents the step of downloading software, makes obtaining disconnectivity data more accessible, and, by that, consequently speeds up scientific progress. However, while the integration of deep learning into various domains holds immense potential, its application still faces constraints. One prevailing limitation, which also pertains to our study, is the issue of generalizability. The efficacy of an algorithm heavily relies on the quality and representativeness of the training data it receives (Taylor & Nitschke, 2018). Consequently, ensuring a diverse and comprehensive dataset becomes crucial for maximizing the model's ability to generalize effectively. Within the realm of medical diagnosis, this challenge is particularly pronounced, exacerbated by biases and stringent privacy policies that restrict data availability as well as typical bias in the distribution of clinical data (Gianfrancesco et al., 2018). However, in the specific context of disconnectomes, we are fortunate to have addressed this limitation by training our model on a broad sample of diverse synthetically generated lesions. Leveraging the capabilities of the BCBToolkit, we can generate an abundance of training data by computing disconnectomes. This unique advantage augments our framework and unlocks novel opportunities for further exploration. The deep-disconnectome is theoretically able to capture the pattern of disconnection for any type of focal lesion (such as focal multiple sclerosis lesions, brain tumors, and post-traumatic lesions). In future work, we can extend the model's training to include alternative focal lesions and examine the correlation between deep-disconnectome outcomes and associated neuropsychological manifestations. Incorporating actual patient data with alternative focal lesions into our deep-learning model would enhance its overall generalizability and establish a more comprehensive foundation for predicting neuropsychological scores following brain damage. Despite the notable time-saving and precise outcomes achieved by AI models, there remains a prevailing mistrust among individuals regarding their reliability in healthcare diagnoses. While statistics demonstrate the better performance of AI models and the number of FDA-approved AI algorithms is steadily growing, concerns stem from the inherent opacity of these "black box" models, which oftentimes do not provide sufficient explanations for their results (Bzdok, Krzywinski & Altman, 2017). This lack of transparency undermines trust in their capabilities. With our framework for predicting stroke outcomes, we mitigate this issue by creating the deep-disconnectomes, which serve as a crucial intermediary, providing a direct connection between the structure and function of the brain and aiding in the interpretation of the predictions rather than predicting neuropsychological scores from lesions directly.