Abstract
Correct identification of anatomical structures is important for safe oncologic thoracic surgery, but inflammation and tissue changes after neoadjuvant therapy can obscure the operative field. Since video-assisted thoracoscopic surgery (VATS) is performed through a camera system, AI-based identification of anatomy can be integrated without altering the surgical setup. The final success in such a task is hindered by scarce expert annotations, target structures covering only a small fraction of the image, and real-time requirements. In addition, AI-based systems rely on network configuration and on pretrained weights that greatly affect the performance of identifying anatomical structure and tissue.
We present FTSpec-UNet, which combines a pretrained Foundation Encoder (ResNet or DINOv2) with a task-specific encoder trained from scratch, allowing the configuration to be adapted to a new procedure without repeating pretraining. On SurgiMind, an in-house dataset of 1539 frames from 81 VATS resections videos, FTSpec-UNet reaches an average Dice of 0.68 across eight anatomical structures, outperforming pretrained U-Net, DeepLabV3+, and SegFormer, and runs in real time on a single consumer GPU. On the public CholecSeg8K benchmark, it provides comparable results to pretrained state-of-the-art baselines. We also propose the Segmentation True Negative Rate (S-TNR), which measures whether absent structures are correctly ignored. This measure complements the background Dice score to provide a better performance picture of AI-based systems, evaluating both segmentation quality and whether absent classes are correctly ignored.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/AMAI_043.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/profile?id=%7EDennis_Barnes1
BibTex
@InProceedings{BarDen_DualEncoder_MICCAISAT2026,
author = { Barnes, Dennis AND Rodriguez-Sanchez, Antonio AND Rodríguez, Ana Lamas AND Ponholzer, Florian},
title = { { Dual-Encoder Transfer Learning for Surgical Image Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17273},
month = {pending},
page = {pending}
}