DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES
M.Tech Research Thesis Colloquium
Speaker : Mr. Ramanujam Narayanan
S.R. Number : 06-18-01-10-22-22-1-21772
Title : “Lesion Synthesis using Physics-Based Noise Models for Low-Data Medical Imaging Regime applications”
Research Supervisor : Dr. Vaanathi Sundaresan
Date & Time : December 11, 2024, 11.00 am
Venue : # 102 CDS Seminar Hall
ABSTRACT
Lesion segmentation and their progression prediction in medical imaging relies critically on the availability of manually annotated, heterogeneous large pathological datasets. Acquiring such diverse large datasets is also challenging because they require coordination and ethical clearances from multiple sites, and the manual annotation of these images is both time-consuming and expensive. On the other hand, this data diversity is difficult to achieve in low-data regimes, hindering the robust training of the models.
Our study presents a lesion simulation method involving structural localized perturbation of healthy tissue using noise models based on the physics of modalities. Later, we localize these perturbations within masks defined by composites of ellipsoidal polygons (thus forming random shapes) and blended them with the input image with varying contrast. The lesion simulation step does not require training and can generate any number of lesions with texture, size, and scale variations, injecting sufficient variability in the training data pool in low-data regimes. We evaluate the performance of our simulated lesions for a downstream lesion segmentation task and show superior performance than its fully supervised counterpart. We also performed extensive ablation studies and experimentally determined the optimal simulation data and minimal training data required for training the segmentation model. Further, we also showed detailed variation charts depicting the possible simulations the method can generate across different datasets. We evaluate the performance on publicly available pathological brain MRI, liver CT, retinal fundus imaging and breast Ultrasound datasets with diverse lesions. Using only 75% of labelled real-world data, the proposed method significantly improves the segmentation performance compared to the fully supervised training, with a 16% mean increase in the Dice score (DSC) and a 33% mean decrease in the 95th percentile of the normalised Hausdorff distance (HD95 (norm)).
We also discuss two potential use cases of the method: (i) for improving segmentation of lacune infarcts by providing simulated instances for augmenting with real lacunar infarct dataset to training lacune segmentation model with improved performance, (ii) for improving the post-operation (post-op) prognosis prediction of glioblastomas using Response Assessment in Neuro-Oncology (RANO) criterion-informed simulations augmenting real post-op and follow-up scans.
ALL ARE WELCOME
