Semantic Segmentation of Medical Images

Summer Internship at Medical Deep Learning and AI Lab (MeDAL) IIT Bombay, Summer 2018

Lung Segmentation from Chest X Rays

[Code]

With the widespread, prevalance of Chest X-Rays (CXRs) in medical diagnosis, CXRs contribute to a majority of the workload of radiologists and other medical practitioners. Thus, it becomes imperative to expedite the analysis of CXRs, with deep-learning being one possible method of automation.

Overview of Algorithm

A variety of Fully Convolutional Networks (Global Convolutional Network, VGG UNet, SegNet, HDC/DUC and Mask R-CNN) were tested to segment the lungs from CXRs. A hybrid loss function, combining Soft Dice Loss, Soft Inverse Dice Loss, and Binary Cross-Entropy Loss (with logits), was employed to achieve the peak performance.

Nuclei Segmentation from Histopathological Images

Overview of Algorithm

I also worked on applying the same models to segment out Nuclei from h/he stained histopathological sections of various organs. The goal was to be able to generalise nuclei detection across various organs. Finetuning of the models along with conventional post-processing techniques enabled us to improve the state of the art results of nuclei segmentation to a F1 score to 85%. This work was used to host the MoNuSeg (Multi-Organ Nuclei-Segmentation) Challenge at MICCAI 2018.

Screening of Precancerous Oral Lesions

Oral cancer is the deadliest form of cancer in India, due to the general poor oral hygiene combined with the lateness of symptoms. A user facing Android app was developed to screen patients and speed up remote diagnosis while reducing the workload of doctors. I annotated and pre-processed pictures of oral lesions from around 800 patients which was later used to train a Siamese Network classifier.

Overview of Algorithm