10 March 2026 in Digestive Endoscopy
Multimodal large language models (MLLMs) can automatically analyze clinical video, but evidence from full esophagogastroduodenoscopy (EGD) and the impact of on‐screen computer‐aided detection/diagnosis (CAD) overlays on MLLM behavior remain unclear. We tested whether an MLLM can produce clinically adequate EGD reports and whether a CAD overlay changes performance. We analyzed five complete EGD videos with Gemini 2.5 Pro in paired versions: (1) clean video and (2) the same video with a CAD overlay. Five blinded endoscopists rated report adequacy in three domains. MLLM accuracy for landmarks/lesions was further assessed by two blinded expert endoscopists using the time‐window rule (a model detection counted as correct if it occurred within ±2 s of the expert‐annotated timestamp). In this retrospective pilot study, five archived diagnostic EGD procedures from five patients were available as full‐length videos. Across five raters, MLLM Completeness was judged adequate in 56.0% (14/25 ratings) with Clean‐Video versus 48.0% (12/25 ratings) with Overlay‐Video ( p = 0.500). Visualization was identical (36.0% [9/25 ratings] for both; p = 1.000). Lesions characteristics were identical (16.0% [4/25] for both; p = 1.00). For the Landmark agreement, the overall accuracy of the MLLM with Clean‐Video vs. Overlay‐Video was: 0.55 [95% CI 0.43–0.67] vs. 0.33 [0.23–0.46], p = 0.029; sensitivity 0.53 [0.40–0.66] vs. 0.35 [0.24–0.49], p = 0.122; specificity 0.67 [0.35–0.88] vs. 0.22 [0.06–0.55], p = 0.125. In this pilot study, Gemini 2.5 Pro demonstrated inadequate performance for clinical EGD reporting. These hypothesis‐generating findings suggest substantial optimization and larger‐scale validation are required before deployment.
21 January 2026 in Computer Methods and Programs in Biomedicine
Accurate assessment of colorectal lesion morphology during colonoscopy is essential for guiding treatment and estimating cancer risk. The Paris classification is widely adopted for this purpose but suffers from substantial inter-observer variability, while Vision Transformers (ViTs) can base their decisions on diffuse, off-lesion attention patterns that are hard to interpret. This study investigates whether directly supervising ViT attention maps with expert lesion annotations can concurrently improve Paris classification performance and model explainability. We propose a Lesion-Focused Attention Loss (LLFA), an attention-supervised pretraining objective that uses expert polyp bounding boxes to focus last-layer [CLS] attention on annotated lesion regions, followed by standard cross-entropy fine-tuning. LLFA is applied to six ViT architectures and evaluated on the public SUN dataset for binary (0-I vs. 0-II) and three-class (0-Ip, 0-Is, 0-IIa) Paris classification. Performance is assessed using frame-wise accuracy and the AttIn, we additionally perform an ablation study against a Grad-CAM consistency baseline. Attention-supervised pretraining yields consistent gains in both accuracy and lesion-focused attention. Across the six ViTs, adding LLFA improves three-class accuracy by up to 7 percentage points. In a detailed ablation on ViT-B/16, LLFA outperforms a Grad-CAM consistency baseline by about 5-13 percentage points across the 2-class and 3-class tasks, and χ2 tests confirm a significant association between high AttIn and correct predictions. Direct supervision of ViT attention with LLFA leverages expert knowledge to jointly boost Paris classification accuracy and spatial interpretability, and compares favourably with Grad-CAM-based explanation regularisation. The source code and dataset splits are publicly available at https://github.com/LucaCarlini/SUNDatasetPretraining.
1 October 2026 in Endoscopy
JEJeanine Wakim El-Khoury
+1
28 July 2026 in Endoscopy International Open
Background and study aim: Optical assessment of deep submucosal invasion (T1b) in T1 colorectal cancer (CRC) is critical for selecting endoscopic versus surgical management. We conducted a systematic review and meta-analysis restricted to studies using a paired comparative design, in which narrow-band imaging (NBI) and magnifying chromoendoscopy (MCE) were applied to the same lesion during the same endoscopic session. Patients and methods: MEDLINE, Embase, and CENTRAL were searched from inception to September 19, 2025. Eligible studies compared NBI and MCE applied to the same lesions during the same endoscopic session, with histopathology as the reference standard and T1b defined as ≥1,000 µm. Two reviewers independently screened studies and extracted data; risk of bias was assessed using QUADAS-2 and QUADAS-C. Pooled sensitivity and specificity and between-modality differences were estimated using a Bayesian bivariate HSROC model. Prespecified subgroup analyses compared real-time diagnosis with image-based evaluation. Results: Ten studies comprising 2,814 lesions were included. Pooled sensitivity and specificity were 75.4% (95% credible interval [CrI] 63.2–85.3) and 94.5% (95% CrI 90.8–97.0) for NBI, and 79.8% (95% CrI 75.9–83.3) and 93.8% (95% CrI 91.6–95.6) for MCE. Direct comparisons indicated small differences in sensitivity and specificity, with uncertainty intervals including no difference. In subgroup analyses, real-time diagnosis showed higher sensitivity than image-based evaluation, while specificity was comparable. Conclusions: NBI and MCE show high accuracy for diagnosing deep submucosal invasion in T1 CRC, with no clear overall superiority in paired comparisons. Diagnostic setting may influence sensitivity and should be considered when translating optical diagnosis studies into practice.
23 September 2026 in Endoscopy