DeepPath Neurodegeneration: ML on IHC Brain Images

DeepPath Icon

Background

Brain tumour research and clinical trials have historically underrepresented minority populations (Taha et al., 2020). Moreover, subtype determination from H&E slides can be challenging for even expert pathologists, delaying care and creating inconsistent results that affect downstream therapy selection (Liu et al., 2022).

We aimed to build a machine learning system that classifies glioma subtypes from histopathology whole-slide images. The database we chose is one of the largest in Asia, with a focus on Indian demographics (Chauhan et al., 2024). Ultimately, we hope to increase throughput of accurate diagnoses, and enable personalized treatment decisions for underserved populations.

Methods

Pre-Processing

484 slides from 328 patients had been diagnosed with one of three glioma subtypes: astrocytoma, oligodendroglioma, and glioblastoma. The pre-processing script takes in whole-slide histochemical images and defines patch sizes (256x256 pixels) to extract from each image, which is passed into ResNet18.

CNN Training and Output

We used the convolution neural network ResNet18, which is pre-trained on the ImageNet dataset. For our purposes, this was used as a backbone, and we fine-tuned it by replacing the final layer with three output nodes which correspond to glioma subtypes. The model’s performance was evaluated using a classification report and confusion matrix, which highlighted overall accuracy.

Goals

We designed our model for use in pathology departments and oncology research centers (Princess Margaret Cancer Centre, UHN, Ontario Institute for Cancer Research), with secondary applications for hospital IT groups and research labs. Key use cases include:

Screening for clinical trials by quickly filtering slides by predicted subtype
Assisting pathologists with slide triage and second reads to improve throughput and consistency
Enabling quality assurance by flagging discordant cases for review and training

Challenges

Scientific limitations were centered around the database itself, which focused on mostly Indian demographics, with limited data coming from other underrepresented populations. Computational limitations resulted from ResNet18, which is relatively shallow. Additionally, ResNet is pretrained on natural images which have different statistics than histopathology patches.

Future Directions

Our project identified promising directions for the use of ML in classifying IHC glioma subtypes. With further time and resources, this project has the potential to shift from a proof-of-concept to a working system that supports research/clinical centres and hospital networks, while serving underrepresented populations.

Citations

Chauhan, E., Sharma, A., Uppin, M. S., Kondamadugu, M., Jawahar, C. V., & Vinod, P. K. (2024). IPD-Brain: An Indian histopathology dataset for glioma subtype classification. Scientific Data, 11(1), 1403. https://doi.org/10.1038/s41597-024-04225-9

Liu, H., Xu, W.-D., Shang, Z.-H., Wang, X.-D., Zhou, H.-Y., Ma, K.-W., Zhou, H., Qi, J.-L., Jiang, J.-R., Tan, L.-L., Zeng, H.-M., Cai, H.-J., Wang, K.-S., & Qian, Y.-L. (2022). Breast Cancer Molecular Subtype Prediction on Pathological Images with Discriminative Patch Selection and Multi-Instance Learning. Frontiers in Oncology, 12. https://doi.org/10.3389/fonc.2022.858453

Taha, B., Winston, G., Tosi, U., Hartley, B., Hoffman, C., Dahmane, N., Mason, C. E., & Greenfield, J. P. (2020). Missing diversity in brain tumor trials. Neuro-Oncology Advances, 2(1), vdaa059. https://doi.org/10.1093/noajnl/vdaa059