Interpretable Machine Learning for Brain Tumour Segmentation & Classification
Brain tumours are the cause for the greatest number of deaths out of all central nervous system related cancers. Still, medical experts have not been able to discover the causes for brain tumours or unique symptoms related to brain tumours. The survival rate of patients with a malignant form of brain tumours like Glioblastoma is as low as 6%-9%. Inability to diagnose and identify their types at early stages further reduces the survival rate of brain tumour patients considerably. Magnetic Resonance Imaging (MRI) and histopathology are the most common diagnostic processes followed by medical officers to identify brain tumours. However, these imaging techniques generate hundreds of images making the diagnostic process exhaustive and hinder early diagnosis and treatments for tumours. This is where machine learning comes into the picture to automate segmentation and classification processes and make the life of radiologists simpler.
Currently there are many systems to classify and segment images that aim to improve the accuracy of their predictions. When it comes to high stake fields like medical diagnosis it is not all about improving the accuracy. The system has to make sense at least for domain experts. This research tries to address the above problem by developing a deep neural network with high accuracy to segment and classify brain tumours using MRI and pathology images that is interpretable (understandable to humans unlike generic black-box models). Hence, it is reliable and can be applied to real-world scenarios instead of existing models and support exhaustive tasks of radiologists. Our main motivation is to provide a complete solution to simplify highly time consuming routine tasks of medical practitioners dealing with a large number of images while replicating the real-world medical diagnostic procedure using multimodal images (MRI & Whole Slide Image (WSI)).
MRI are subjected to both segmentation and classification while WSI is only classified into one of 3 classes: Glioblastoma, Oligodendroglioma or Astrocytoma. Variational auto-encoder with 3D U-Net is used for segmentation of MRI as it overcomes the gradient vanishing problem through skip connections and variational auto-encoder branch reconstructs the input image itself to regularize the shared decoder and impose additional constraints on its layers. Two DenseNet-BC models are used to classify MRI volumes and WS images and predict output as probabilities for each class. Then the outputs from both the models are ensembled to predict the final class label. Gradient Weighted Class Activation Map (Grad-CAM) is used as the interpretable approach to highlight the localization of WSI and MRI which interpret the attentions of neural networks in classification.
More:
Interpretable Machine Learning for Brain Tumor Analysis Using MRI
Interpretable Machine Learning for Brain Tumour Analysis using MRI and Whole Slide Images
Datasets:
MRI segmentation Medical Segmentation Decathlon - Task01_BrainTumor
MRI & WSI classification MICCAI 2020 Combined Radiology and Pathology Classification Challenge
National cancer institute Genomic Data Commons (GDC) Data Portal
Cancer imaging archive Histopathology imaging on TCIA