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:124@cds.iisc.ac.in DTSTART;TZID=Asia/Kolkata:20250526T143000 DTEND;TZID=Asia/Kolkata:20250526T153000 DTSTAMP:20250522T111857Z URL:https://cds.iisc.ac.in/events/ph-d-thesis-colloquium-cds-may-26-2025-u nsupervised-test-time-adaptation-for-patient-specific-deep-learning-models -in-medical-imaging/ SUMMARY:Ph.D. Thesis Colloquium: CDS: May 26\, 2025 "Unsupervised test-time adaptation for patient-specific deep learning models in medical imaging" DESCRIPTION:\n\nDEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES\nPh.D.  (ERP ) Thesis Colloquium \n\n\n\n\n\nSpeaker                 : Mr. H ariharan Ravishankar (ERP Ph.D. candidate)\nS.R. Number        : 06-18 -00-11-12-20-1-18917\n\nTitle                        :Unsuperv ised test-time adaptation for patient-specific deep learning models in med ical imaging\nResearch Supervisors:  Phaneendra Kumar Yalavarthy &\; D r. Prasad Sudhakar (GE HealthCare)\nDate &\; Time         : May 2 6\, 2025 (Monday) at 2:30 PM\nVenue                     : # 102 CDS Seminar Hall\n\n\n\n\n\n\n\nABSTRACT\n\nDeep learning (DL) models hav e achieved state-of-the-art results in multiple medical imaging applicatio ns\, resulting in the widespread adoption of artificial intelligence (AI) models for radiological workflows. Despite their success\, they exhibit tw o major weaknesses in real-world applications: a) poor generalization\, an d b) lack of patient optimality. The potential `distribution shift' in uns een medical imaging data\, due to changes in demography\, acquisition hard ware\, etc.\, often leads to a reduction in performance. Solving this prob lem is critical in medical image analysis because this variance in perform ance across cases and abject failures on selected subjects will increase t he burden on care-giving experts and reduce trust in AI-based applications . The aim of this thesis is to develop methods for patient-specific test-t ime adaptation of pretrained DL models on individual subject data. While g eneralization is often treated as a training time goal\, test-time adaptat ion (TTA) modifies the weights during inference time in an unsupervised ma nner specific to each subject’s data to jointly address generalization a nd patient optimality. While sustaining performance in individual subjects is seen as a challenge for current DL models\, this line of research also offers an opportunity for next-generation clinical decision support syste ms. AI methods that would result in patient-specific DL models can be hera lded as the next big step in medical imaging.\n\nInformation Geometric Tes t Time Adaptation (IGTTA) for semantic segmentation:  The test-time adapt ation (TTA) of deep-learning-based semantic segmentation models\, specific to individual patient data\, was addressed in this part of thesis work. E xisting TTA methods in medical imaging are often unconstrained and require prior anatomical information or additional neural networks built during t he training phase\, making them less practical and prone to performance de terioration. In this part of thesis work\, a novel framework based on info rmation geometric principles was proposed to achieve generic\, off-the-she lf\, regularized patient-specific adaptation of models during test-time. B y considering the pre-trained model and adapted models as part of statisti cal neuromanifolds\, test-time adaptation was treated as constrained funct ional regularization using information geometric measures\, leading to imp roved generalization and patient optimality. The efficacy of the proposed approach was shown for three challenging problems: a) improving the genera lization of state-of-the-art models for segmenting COVID-19 anomalies in C omputed Tomography (CT) images\, b) cross-institutional brain tumor segmen tation from magnetic resonance (MR) images\, and c) segmentation of retina l layers in Optical Coherence Tomography (OCT) images. Furthermore\, it wa s demonstrated that robust patient-specific adaptation can be achieved wit hout adding a significant computational burden\, making it the first of it s kind\, based on information geometric principles.\n\nInference-time adap tation via domain transforms (ITADT): OCT imaging has emerged as the modal ity of choice for the diagnosis of retinal diseases. The popular\, cost-ef fective variant of OCT - Spectral-Domain Optical Coherence Tomography (SDO CT) often suffers in image quality owing to speckle noise\, making expert or automated analysis of these images challenging. This part of thesis wor k proposes for the first-time inference time adaptation of deep learning m odels for a given test sample to achieve improved diagnostic performance. The adaptation is driven by domain transforms\, a framework in which domai n-specific augmentation is employed for a test sample\, and the deep learn ing model weights are updated in a constrained and regularized manner. The performance of the proposed approach was evaluated on both simulated and real-world noisy B-mode OCT images affected by varying degrees of speckle noise from four benchmark SD-OCT datasets. Systematic studies  clearly de monstrate the utility of the proposed approach in building generalized and robust deep learning models for automated retinal disease diagnosis using OCT images.\n\nUncertainty-aware test-time adaptation for inverse problem s (UATTA): Deep learning models for inverse problems (both one-pass and un rolled versions): have these limitations: 1) inability to provide pixel-wi se uncertainty quantification\; 2) sensitivity to variation in forward mod el\; and 3) suboptimal regularization over different iterations. In this p art of thesis work\, uncertainty aware test-time adaptation of off-the-she lf models is proposed for improved image reconstruction performance. Diffe rent pixel-wise uncertainty quantification methods were explored for their utility in test-time adaptation\, along with data fidelity constraints. T he results were studied to address the pertinent problem of quantitative s usceptibility mapping for quantifying tissue magnetic susceptibility. The proposed framework provides reliable and accurate reconstruction of these images.\n\nTest-time adaptation with foundational models (TTAFM): Recently \, foundational models have demonstrated impressive promptable segmentatio n of natural images. However\, these models often perform poorly with medi cal imaging data. In this part of the thesis\, a two-way synergic approach for jointly utilizing foundational models while adapting pretrained medic al imaging segmentation models is explored. The results are demonstrated o n various medical imaging semantic segmentation tasks to show the ability of foundational models as neural surrogates\, thus demonstrating the test- time adaptation achieved in real time. As these foundational models are ca pable of handling high levels of semantics\, using them as surrogates help s reduce reliance on costly annotated data needed to achieve state-of-the- art results.\n\nIn summary\, this thesis work developed methods to improve the generalization of typical deep learning models utilized in medical im aging by personalizing them to individual patients during testing in an un supervised and automated manner. The developed methods demonstrate the app licability and utility of different medical imaging tasks across various m odalities. This thesis work is the first of its kind in medical imaging to showcase the utility of test-time adaptation methods\, paving the way for further research towards the goal of precision medicine.\n\n\n\n\n\nALL A RE WELCOME\n\n\n\n \; CATEGORIES:Events,Ph.D. Thesis Colloquium END:VEVENT BEGIN:VTIMEZONE TZID:Asia/Kolkata X-LIC-LOCATION:Asia/Kolkata BEGIN:STANDARD DTSTART:20240526T143000 TZOFFSETFROM:+0530 TZOFFSETTO:+0530 TZNAME:IST END:STANDARD END:VTIMEZONE END:VCALENDAR