MICCAI 2026 Accepted · Video Prostate Segmentation · Temporal Consistency · Ultrasound

Distilling Temporal Coherence into 2D Networks for TRUS Prostate Video Segmentation

Dong Yeong Kim^*, JunGyu Lee^*†, Jaewon Choi, June Young Seo, Myeongseop Kim, Jinwook Choi, Taek Min Kim^‡, Young-Gon Kim^‡ MICCAI 2026 · * Equal contribution · † Work done prior to joining Yonsei Univ. · ‡ Corresponding authors

DTC-TRUS trains a standard 2D segmentation network to behave consistently across TRUS video frames, while preserving efficient single-frame inference at test time.

Paper (arXiv) Code Dataset

TRUS-V: 2,679 frames
Views: Axial + Sagittal
Inference: 2D only

Teaser comparing flickering 2D slice-based predictions with robust temporally consistent predictions. — Teaser. Slice-based 2D segmentation suffers from inter-frame flickering, whereas DTC-TRUS enforces temporal coherence during training.

Abstract

Temporal stability without temporal inference overhead.

Real-time prostate segmentation in transrectal ultrasound (TRUS) video is challenging because conventional 2D models ignore temporal context and produce flickering masks, whereas 3D or recurrent video models increase inference latency. DTC-TRUS addresses this trade-off by distilling temporal coherence into a 2D student network during training.

The framework combines confidence-weighted optical-flow consistency, dual-scale prototype alignment, self-supervised geometric equivariance, and knowledge distillation from a static-image teacher. At inference, only the 2D student is used, avoiding optical-flow computation, teacher inference, or temporal modules.

Motivation

Why temporal coherence matters in TRUS video.

2D slice-based models

Efficient frame-by-frame inference, but predictions can fluctuate across adjacent frames because temporal context is ignored.

3D / recurrent models

Can model spatiotemporal structure, but add computation that is difficult to justify for low-latency intra-operative use.

DTC-TRUS

Uses video dynamics only as training supervision, so the deployed model remains a standard 2D segmenter.

Method

Temporally Consistent Learning Framework

DTC-TRUS supervises a student network using pixel-level motion consistency and feature-level semantic alignment, while retaining anatomical priors from a frozen teacher.

Overview of DTC-TRUS with student-teacher learning, pseudo-labeling, optical flow consistency, and prototype alignment. — Framework overview. The training objective combines self-supervised segmentation, knowledge distillation, confidence-weighted temporal consistency, and prototype alignment. Only the 2D student is used at inference.

Self-supervised equivariance

Pseudo-labels are generated from geometrically transformed frames and restored via the inverse transform, encouraging transformation-consistent segmentation without dense per-frame video labels.

Knowledge distillation

A frozen static-image teacher regularizes the student so video adaptation does not erase spatial and anatomical priors learned from labeled images.

Confidence-weighted temporal consistency

Adjacent-frame predictions are aligned by optical-flow warping. A non-occlusion confidence map down-weights unstable regions where acoustic artifacts or motion make the temporal signal unreliable.

Dual-scale prototype alignment

Local foreground prototypes stabilize boundary features, while global background prototypes stabilize scene semantics. Area-adaptive weighting balances these two alignment signals.

Training objective

L_total = λ_seg L_seg + λ_KD L_KD + λ_con L_con + λ_proto L_proto

The deployment path removes all training-only components: no teacher, no optical flow, and no temporal interaction module are required for inference.

Dataset

TRUS-V benchmark

TRUS-V is a multi-view prostate ultrasound video benchmark designed to evaluate segmentation accuracy and temporal coherence in clinically realistic videos.

2,679 annotated frames

10 patients

20 continuous videos

2 views: axial and sagittal

Patient-level split

The benchmark is partitioned at the patient level into 2,400 training frames and 279 testing frames.

Reference masks

Candidate masks are initialized with an ensemble U-Net trained on a separate static dataset and then manually refined frame-by-frame by experienced radiologists.

Release status. TRUS-V is publicly available through the Korea Health Data Platform (KHDP). Access TRUS-V on KHDP

Results

Reported quantitative performance

The tables below summarize the manuscript-reported performance of DTC-TRUS on SUN-SEG and TRUS-V. Values should be updated if the paper changes during review.

SUN-SEG unseen splits

Split	S_α	E^mn_ϕ	F^w_β	F^mn_β	Dice	Sen
Easy	0.816	0.882	0.738	0.784	0.746	0.719
Hard	0.816	0.878	0.719	0.758	0.737	0.741

Reported real-time inference: 89.95 FPS with ACSNet.

TRUS-V benchmark

Method	S_α	E^mn_ϕ	F^w_β	F^mn_β	Dice	Sen
Ours	0.967	0.988	0.830	0.828	0.829	0.839

Reported real-time inference: 127.97 FPS with U-Net++.

Visual comparison on TRUS-V across frames, ground truth, competing methods, and DTC-TRUS results. — Visual comparison on TRUS-V. DTC-TRUS maintains smoother prostate boundaries under acoustic shadows and rapid probe motion.

Materials

Code, paper, and dataset links

Paper

Accepted at MICCAI 2026. The arXiv preprint is available online.

arXiv:2606.31198

Available

Code

Training and inference code are available in the official GitHub repository.

github.com/DYDevelop/DTC-TRUS

Available

Dataset

The TRUS-V benchmark is available through the Korea Health Data Platform (KHDP).

khdp.net/database/data-search-detail/TRUS-V

Available

Citation

arXiv preprint is available now. Proceedings details and DOI will be added after publication.

@inproceedings{kim2026dtctrus,
  title     = {Distilling Temporal Coherence into 2D Networks for Transrectal Ultrasound Prostate Video Segmentation},
  author    = {Kim, Dong Yeong and Lee, JunGyu and Choi, Jaewon and Seo, June Young and Kim, Myeongseop and Choi, Jinwook and Kim, Taek Min and Kim, Young-Gon},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI)},
  year      = {2026},
  note      = {Accepted; arXiv:2606.31198},
  eprint    = {2606.31198},
  archivePrefix = {arXiv},
  primaryClass = {cs.CV},
  url       = {https://arxiv.org/abs/2606.31198}
}