BEGIN:VCALENDAR VERSION:2.0 PRODID:-//wp-events-plugin.com//7.2.3.1//EN TZID:Asia/Kolkata X-WR-TIMEZONE:Asia/Kolkata BEGIN:VEVENT UID:145@cds.iisc.ac.in DTSTART;TZID=Asia/Kolkata:20250926T153000 DTEND;TZID=Asia/Kolkata:20250926T163000 DTSTAMP:20250911T074559Z URL:https://cds.iisc.ac.in/events/cds-phd-thesis-defense-sep-26-friday-330 -pm-unsupervised-test-time-adaptation-for-patient-specific-deep-learning-m odels-in-medical-imaging/ SUMMARY:CDS PhD Thesis Defense: Sep-26 (Friday) @ 3:30 PM: "Unsupervised te st-time adaptation for patient-specific deep learning models in medical im aging" DESCRIPTION:DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES\nPh.D. (ERP) Thes is Defense\n\n\n\nSpeaker : Mr. Hariharan Ravishankar (ERP Ph.D. candidate )\nS.R. Number : 06-18-00-11-12-20-1-18917\nTitle :Unsupervised test-time adaptation for patient-specific deep learning models in medical imaging\nR esearch Supervisors: Phaneendra K. Yalavarthy &\; Dr. Prasad Sudhakar ( GE HealthCare)\nThesis Examiner : Prof. Debdoot Sheet\, IIT-Kharagpur\nDat e &\; Time : Sep 26\, 2025 (Friday) at 3:30 PM\nVenue : # 102 CDS Semin ar Hall\n\n\n\nABSTRACT\nDeep learning (DL) models have achieved state-of- the-art results in multiple medical imaging applications\, resulting in th e widespread adoption of artificial intelligence (AI) models for radiologi cal workflows. Despite their success\, they exhibit two major weaknesses i n real-world applications: a) poor generalization\, and b) lack of patient optimality. The potential `distribution shift' in unseen medical imaging data\, due to changes in demography\, acquisition hardware\, etc.\, often leads to a reduction in performance. Solving this problem is critical in m edical image analysis because this variance in performance across cases an d abject failures on selected subjects will increase the burden on care-gi ving experts and reduce trust in AI-based applications. The aim of this th esis is to develop methods for patient-specific test-time adaptation of pr etrained DL models on individual subject data. While generalization is oft en treated as a training time goal\, test-time adaptation (TTA) modifies t he weights during inference time in an unsupervised manner specific to eac h subject’s data to jointly address generalization and patient optimalit y. While sustaining performance in individual subjects is seen as a challe nge for current DL models\, this line of research also offers an opportuni ty for next-generation clinical decision support systems. AI methods that would result in patient-specific DL models can be heralded as the next big step in medical imaging.\n\nInformation Geometric Test Time Adaptation (I GTTA) for semantic segmentation: The test-time adaptation (TTA) of deep-le arning-based semantic segmentation models\, specific to individual patient data\, was addressed in this part of thesis work. Existing TTA methods in medical imaging are often unconstrained and require prior anatomical info rmation or additional neural networks built during the training phase\, ma king them less practical and prone to performance deterioration. In this p art of thesis work\, a novel framework based on information geometric prin ciples was proposed to achieve generic\, off-the-shelf\, regularized patie nt-specific adaptation of models during test-time. By considering the pre- trained model and adapted models as part of statistical neuromanifolds\, t est-time adaptation was treated as constrained functional regularization u sing information geometric measures\, leading to improved generalization a nd patient optimality. The efficacy of the proposed approach was shown for three challenging problems: a) improving the generalization of state-of-t he-art models for segmenting COVID-19 anomalies in Computed Tomography (CT ) images\, b) cross-institutional brain tumor segmentation from magnetic r esonance (MR) images\, and c) segmentation of retinal layers in Optical Co herence Tomography (OCT) images. Furthermore\, it was demonstrated that ro bust patient-specific adaptation can be achieved without adding a signific ant computational burden\, making it the first of its kind\, based on info rmation geometric principles.\n\nInference-time adaptation via domain tran sforms (ITADT): OCT imaging has emerged as the modality of choice for the diagnosis of retinal diseases. The popular\, cost-effective variant of OCT - Spectral-Domain Optical Coherence Tomography (SDOCT) often suffers in i mage quality owing to speckle noise\, making expert or automated analysis of these images challenging. This part of thesis work proposes for the fir st-time inference time adaptation of deep learning models for a given test sample to achieve improved diagnostic performance. The adaptation is driv en by domain transforms\, a framework in which domain-specific augmentatio n is employed for a test sample\, and the deep learning model weights are updated in a constrained and regularized manner. The performance of the pr oposed approach was evaluated on both simulated and real-world noisy B-mod e OCT images affected by varying degrees of speckle noise from four benchm ark SD-OCT datasets. Systematic studies clearly demonstrate the utility of the proposed approach in building generalized and robust deep learning mo dels for automated retinal disease diagnosis using OCT images.\n\nUncertai nty-aware test-time adaptation for inverse problems (UATTA): Deep learning models for inverse problems (both one-pass and unrolled versions): have t hese limitations: 1) inability to provide pixel-wise uncertainty quantific ation\; 2) sensitivity to variation in forward model\; and 3) suboptimal r egularization over different iterations. In this part of thesis work\, unc ertainty aware test-time adaptation of off-the-shelf models is proposed fo r improved image reconstruction performance. Different pixel-wise uncertai nty quantification methods were explored for their utility in test-time ad aptation\, along with data fidelity constraints. The results were studied to address the pertinent problem of quantitative susceptibility mapping fo r quantifying tissue magnetic susceptibility. The proposed framework provi des reliable and accurate reconstruction of these images.\n\nTest-time ada ptation with foundational models (TTAFM): Recently\, foundational models h ave demonstrated impressive promptable segmentation of natural images. How ever\, these models often perform poorly with medical imaging data. In thi s part of the thesis\, a two-way synergic approach for jointly utilizing f oundational models while adapting pretrained medical imaging segmentation models is explored. The results are demonstrated on various medical imagin g semantic segmentation tasks to show the ability of foundational models a s neural surrogates\, thus demonstrating the test-time adaptation achieved in real time. As these foundational models are capable of handling high l evels of semantics\, using them as surrogates helps reduce reliance on cos tly annotated data needed to achieve state-of-the-art results.\n\nIn summa ry\, this thesis work developed methods to improve the generalization of t ypical deep learning models utilized in medical imaging by personalizing t hem to individual patients during testing in an unsupervised and automated manner. The developed methods demonstrate the applicability and utility o f different medical imaging tasks across various modalities. This thesis w ork is the first of its kind in medical imaging to showcase the utility of test-time adaptation methods\, paving the way for further research toward s the goal of precision medicine.\n\n\n\nALL ARE WELCOME CATEGORIES:Events,Thesis Defense END:VEVENT BEGIN:VTIMEZONE TZID:Asia/Kolkata X-LIC-LOCATION:Asia/Kolkata BEGIN:STANDARD DTSTART:20240926T153000 TZOFFSETFROM:+0530 TZOFFSETTO:+0530 TZNAME:IST END:STANDARD END:VTIMEZONE END:VCALENDAR