MUMBAI, India, June 30 -- Intellectual Property India has published a patent application (202641057462 A) filed by Madhu Shree Aravindan; Dr. S. Ramamoorthy; and Dr, P. Madhavan on May 06, 2026, for Consistency Aware Medical Image Segmentation Using Medsam-3.
Inventors include Madhu Shree Aravindan; Dr. S. Ramamoorthy; and Dr, P. Madhavan.
The application for the patent was published on June 26, 2026, under issue no. 26/2026.
Abstract: Vision foundation models like SAM-3 exhibit strong zero-shot 2D performance but struggle with spatial coherence in 3D volumetric data, often producing "staircase" artifacts. We investigate the trade-offs between computational efficiency and 3D consistency when adapting these models to MRI and CT datasets. We first present a negative result: CrossPlane Contrastive Learning (CPCL) enforces orthogonal consistency but incurs prohibitive costs, requiring over 40 GB VRAM and increasing latency 4.5-fold. To resolve this, we propose Spatial Continuity Regularization (SCR-SAM), a lightweight longitudinal strategy optimizing adjacent slice pairs. Experiments on BraTS and KiTS datasets show SCR-SAM reduces Average Symmetric Surface Distance (ASSD) by approximately 79% on MRI volumes while maintaining a peak training VRAM of only 12.70 GB on Tesla T4 hardware. By leveraging 2D encoder power without 3D architectural overhead, SCR-SAM offers a hardware-accessible pathway for high-fidelity volumetric segmentation. Code is available at: https://github.eom/ellowOrld/SCR-SAM.r
Disclaimer: Curated by HT Syndication.
