Data Efficiency of Segment Anything Model for Optic Disc and Cup Segmentation

Fabian Yii*, Tom MacGillivray, Miguel O. Bernabeu

*Corresponding author for this work

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Abstract / Description of output

We investigated the performance of Segment Anything Model (SAM)—the first promptable foundation model for image segmentation—for optic disc (OD) and optic cup (OC) segmentation when fine-tuned on progressively smaller number of fundus images. Three different implementations of SAM with an input prompt were considered: (1) SAM with an OD/OC-centred bounding box (SAM GT); (2) SAM with a noise-added (e.g. displacement, size variation) bounding box (SAM Noise); and (3) SAM with an automatically predicted (using Faster R-CNN) bounding box (SAM Auto). Two popular pre-trained semantic segmentation models, DeepLabV3 with a MobileNetV3-Large backbone and DeepLabV3 with a ResNet-50 backbone were used as baseline models. For OD segmentation, ResNet-50 exhibited comparable if not higher data efficiency (i.e. good performance despite limited training data) than even the most optimal implementation of SAM (SAM GT), although SAM was evidently more robust to small training set sizes, e.g. 25, than MobileNetV3-Large and in eyes with more challenging OD morphologies, e.g. significant peri-papillary atrophy. For OC segmentation, however, SAM GT and SAM Noise consistently demonstrated higher data efficiency, particularly in eyes with relatively small cup-to-disc ratio and ill-defined OC margin.
Original languageEnglish
Title of host publicationMedical Image Computing and Computer Assisted Intervention – MICCAI 2023 Workshops
Subtitle of host publicationMTSAIL 2023, LEAF 2023, AI4Treat 2023, MMMI 2023, REMIA 2023, Held in Conjunction with MICCAI 2023, Vancouver, BC, Canada, October 8–12, 2023, Proceedings
EditorsJonghye Woo, Alessa Hering, Wilson Silva, Xiang Li, Huazhu Fu, Xiaofeng Liu, Fangxu Xing, Sanjay Purushotham, Tejas S. Mathai, Pritam Mukherjee, Max De Grauw, Regina Beets Tan, Valentina Corbetta, Elmar Kotter, Mauricio Reyes, Christian F. Baumgartner, Quanzheng Li, Richard Leahy, Bin Dong, Hao Chen, Yuankai Huo, Jinglei Lv, Xinxing Xu, Xiaomeng Li, Dwarikanath Mahapatra, Li Cheng, Caroline Petitjean, Benoît Presles
PublisherSpringer, Cham
Pages336-346
Number of pages11
Volume14394
ISBN (Electronic)978-3-031-47425-5
ISBN (Print)978-3-031-47424-8
DOIs
Publication statusPublished - 3 Feb 2024
Event26th International Conference on Medical Image Computing and Computer Assisted Intervention, 2023 - Vancouver, Canada
Duration: 8 Oct 202312 Oct 2023
Conference number: 26

Publication series

NameLecture Notes in Computer Science
PublisherSpringer
Volume14394
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference26th International Conference on Medical Image Computing and Computer Assisted Intervention, 2023
Abbreviated titleMICCAI 2023
Country/TerritoryCanada
CityVancouver
Period8/10/2312/10/23

Keywords / Materials (for Non-textual outputs)

  • Segment Anything Model
  • Optic Disc
  • Optic Cup

Fingerprint

Dive into the research topics of 'Data Efficiency of Segment Anything Model for Optic Disc and Cup Segmentation'. Together they form a unique fingerprint.

Cite this