Deterministic and Gaussian Decoder LoRA for Point-Prompted Abdominal CT Segmentation

This study examines deterministic low-rank adaptation (LoRA) and Gaussian decoder LoRA with heuristic rank selection for point-prompted abdominal CT segmentation using SAM2. Both approaches were evaluated on a single BTCV split comprising 19 training, five validation, and six test volumes, with one training seed and 2,466 matched positive-point prompts. Across 78 volume–organ pairs, mean Dice scores were 0.5123 for frozen SAM2, 0.8049 for deterministic LoRA, and 0.7982 for Gaussian decoder LoRA. The deterministic model achieved a 0.667-percentage-point advantage under equal volume– organ weighting (descriptive 95% t interval: −0.214 to +1.549), although the difference became −0.019 points under equal prompt weighting. We also investigated whether decoder uncertainty provides additional information for identifying segmentation failures. Adding full-image mutual information (MI) to a failureranking model containing organ identity, predicted area, and entropy produced a negligible change in mean AUROC (−0.0001). Retrospective inference-only analyses showed similarly limited effects from replacing learned decoder standard deviations with fixed values. Adding dropout MI to organ, area, and entropy on 189 deterministic-model failures changed AUROC from 0.8222 to 0.8238. The Gaussian decoder also required substantially more inference time than its deterministic counterpart. Both adaptation strategies improved overlap relative to frozen SAM2, but the available evidence does not establish a consistent advantage for Gaussian decoder adaptation or its uncertainty estimates beyond simpler alternatives.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-18
DOI
https://doi.org/10.21203/rs.3.rs-11759900/v1
Primary Topic
Medical Image Segmentation Techniques
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Deterministic and Gaussian Decoder LoRA for Point-Prompted Abdominal CT Segmentation

Aritrik Ghosh
6 links
Research Square
Medical Image Segmentation Techniques
preprint

Deterministic and Gaussian Decoder LoRA for Point-Prompted Abdominal CT Segmentation

Aritrik Ghosh
preprint en

Abstract

This study examines deterministic low-rank adaptation (LoRA) and Gaussian decoder LoRA with heuristic rank selection for point-prompted abdominal CT segmentation using SAM2. Both approaches were evaluated on a single BTCV split comprising 19 training, five validation, and six test volumes, with one training seed and 2,466 matched positive-point prompts. Across 78 volume–organ pairs, mean Dice scores were 0.5123 for frozen SAM2, 0.8049 for deterministic LoRA, and 0.7982 for Gaussian decoder LoRA. The deterministic model achieved a 0.667-percentage-point advantage under equal volume– organ weighting (descriptive 95% t interval: −0.214 to +1.549), although the difference became −0.019 points under equal prompt weighting. We also investigated whether decoder uncertainty provides additional information for identifying segmentation failures. Adding full-image mutual information (MI) to a failureranking model containing organ identity, predicted area, and entropy produced a negligible change in mean AUROC (−0.0001). Retrospective inference-only analyses showed similarly limited effects from replacing learned decoder standard deviations with fixed values. Adding dropout MI to organ, area, and entropy on 189 deterministic-model failures changed AUROC from 0.8222 to 0.8238. The Gaussian decoder also required substantially more inference time than its deterministic counterpart. Both adaptation strategies improved overlap relative to frozen SAM2, but the available evidence does not establish a consistent advantage for Gaussian decoder adaptation or its uncertainty estimates beyond simpler alternatives.

Research Square
Swami Vivekananda University (IN)
Good health and well-being
Medical Image Segmentation Techniques
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon