[Paper Review] Assessing generalisability of deep learning-based polyp detection and segmentation methods through a computer vision challenge

Sharib Ali, Noha Ghatwary+30 more|arXiv (Cornell University)|Feb 24, 2022

Colorectal Cancer Screening and Detection19 citations

TL;DR

This study evaluates the generalizability of deep learning-based polyp detection and segmentation models through the EndoCV 2021 challenge, using a multi-center, multi-modality colonoscopy dataset. Top-performing teams achieved over 80% Dice score on diverse validation sets, but the results reveal critical limitations in robustness across unseen centers and imaging conditions, highlighting the urgent need for improved generalization in clinical AI systems.

ABSTRACT

Polyps are well-known cancer precursors identified by colonoscopy. However, variability in their size, location, and surface largely affect identification, localisation, and characterisation. Moreover, colonoscopic surveillance and removal of polyps (referred to as polypectomy ) are highly operator-dependent procedures. There exist a high missed detection rate and incomplete removal of colonic polyps due to their variable nature, the difficulties to delineate the abnormality, the high recurrence rates, and the anatomical topography of the colon. There have been several developments in realising automated methods for both detection and segmentation of these polyps using machine learning. However, the major drawback in most of these methods is their ability to generalise to out-of-sample unseen datasets that come from different centres, modalities and acquisition systems. To test this hypothesis rigorously we curated a multi-centre and multi-population dataset acquired from multiple colonoscopy systems and challenged teams comprising machine learning experts to develop robust automated detection and segmentation methods as part of our crowd-sourcing Endoscopic computer vision challenge (EndoCV) 2021. In this paper, we analyse the detection results of the four top (among seven) teams and the segmentation results of the five top teams (among 16). Our analyses demonstrate that the top-ranking teams concentrated on accuracy (i.e., accuracy > 80% on overall Dice score on different validation sets) over real-time performance required for clinical applicability. We further dissect the methods and provide an experiment-based hypothesis that reveals the need for improved generalisability to tackle diversity present in multi-centre datasets.

Motivation & Objective

• To rigorously evaluate the generalization performance of deep learning models in polyp detection and segmentation across diverse clinical settings.
• To identify the limitations of state-of-the-art models when applied to out-of-distribution data from different endoscopy centers, systems, and imaging modalities.
• To assess whether high accuracy on validation sets translates to reliable performance on unseen, real-world clinical data.
• To provide actionable insights for improving model robustness in clinical endoscopic AI applications.
• To benchmark and analyze the top-performing methods from a large-scale, multi-institutional computer vision challenge.

Proposed method

• Curated a multi-center, multi-population colonoscopy dataset comprising 3,762 annotated frames from 6 different endoscopy centers, including WLE and NBI modalities.
• Organized the EndoCV 2021 challenge to crowdsource development of polyp detection and segmentation models from global machine learning teams.
• Evaluated models on four distinct test sets: NBI-single, WLE-single, seen sequence, and unseen sequence data, representing diverse clinical conditions.
• Used standard metrics including Dice Score (DSC), Jaccard Coefficient (JC), F2-score, PPV, Recall, Accuracy (ACC), and Hausdorff Distance (H d) for quantitative assessment.
• Selected top four teams for detection and top five for segmentation based on performance across validation sets.
• Conducted ablation and methodological analysis to identify design patterns and generalization failures in top models.

Experimental results

Research questions

• RQ1 How well do state-of-the-art deep learning models for polyp detection and segmentation generalize across different endoscopy centers and imaging modalities?
• RQ2 To what extent does high performance on in-distribution validation sets predict success on out-of-distribution, real-world test data?
• RQ3 What architectural or training choices contribute most to robustness or failure in cross-center generalization?
• RQ4 Are there identifiable design patterns in top-performing models that enhance generalization, and how do they compare to baseline approaches?
• RQ5 What are the key failure modes of current models when tested on unseen sequence data and diverse imaging conditions?

Key findings

• The top-performing teams achieved over 80% Dice score on all validation sets, indicating strong in-distribution performance.
• Despite high accuracy, models showed significant performance drops on unseen sequence data (e.g., Dice score dropped to ~0.30–0.40), indicating poor out-of-distribution generalization.
• The best segmentation models achieved a Dice score of 0.6744 ± 0.3011 on seen sequences and 0.4096 ± 0.3577 on unseen sequences, revealing a major drop in robustness.
• Teams prioritized accuracy over inference speed, with no model achieving real-time performance, limiting clinical deployability.
• Analysis revealed that models relying on multi-scale features and attention mechanisms showed better generalization, but still failed on unseen centers and modalities.
• The study identifies a critical gap in current deep learning methods: high performance on curated benchmarks does not ensure reliability in real-world, multi-center clinical deployment.