Presentation Dates

MIUA 2025
15–17 July, University of Leeds (UK)
MICCAI 2025
23–27 September, Daejeon Convention Center (South Korea)

Paper 1:

NERO: Explainable Out-of-Distribution Detection with Neuron-level Relevance
Anju Chhetri, Jari Korhonen, Prashnna Gyawali, Binod Bhattarai

MICCAI 2025

See full paper: arXiv

Deep learning models in medical imaging can fail silently when faced with unfamiliar or out-of-distribution (OOD) inputs, a critical concern in clinical settings. This research introduces NERO, a novel method for OOD detection that focuses on neuron-level relevance patterns rather than high-level features or logits.

By clustering relevance maps for known classes and measuring how far a new sample deviates from these clusters, NERO not only improves detection accuracy but also offers explainable outputs. Tested on gastrointestinal datasets (Kvasir, GastroVision), NERO consistently outperformed existing methods across model architectures.

Paper 2:

NCDD: Nearest Centroid Distance Deficit for Out‑of‑Distribution Detection in Gastrointestinal Vision
Sandesh Pokhrel, Sanjay Bhandari, Sharib Ali, Tryphon Lambrou, Anh Nguyen, Yash Raj Shrestha, Angus Watson, Danail Stoyanov, Prashnna Gyawali, Binod Bhattarai

MIUA 2025 Best Paper Award

See full paper: arXiv

Reliable deep learning in medical imaging requires the ability to flag unfamiliar or anomalous inputs. This challenge is particularly acute in gastrointestinal imaging, where in-distribution and out-of-distribution (OOD) examples often share similar visual features. NCDD frames anomaly detection as an OOD problem and proposes a simple yet effective solution: compute how far a new sample’s feature representation deviates from its nearest class centroid. In-distribution samples cluster close to class centroids, while OOD samples tend to lie farther away. Evaluated on Kvasir2 and GastroVision datasets across different architectures, NCDD consistently outperformed state-of-the-art methods—demonstrating a more reliable way to flag anomalies in medical images.

Paper 3:

Multimodal Federated Learning With Missing Modalities through Feature Imputation Network
Pranav Poudel, Aavash Chhetri, Prashnna Gyawali, Georgios Leontidis, Binod Bhattarai

See full paper: arXiv

Federated learning enables multi-institutional collaboration without sharing raw data, but in healthcare settings, missing data modalities (like uncollected scans or tests) are common challenges. This paper introduces a lightweight feature imputation network to reconstruct missing modality data at the feature level instead of synthesizing raw inputs. Tested across three major chest X-ray datasets (MIMIC‑CXR, NIH Open‑I, CheXpert), both in uniform and varied data conditions, this method improved performance over standard baselines. The approach is efficient, preserves privacy, and supports real-world clinical AI deployment even when data is incomplete.

Paper 4:

Hallucination-Aware Multimodal Benchmark for Gastrointestinal Image Analysis with Large Vision‑Language Models
Bidur Khanal, Sandesh Pokhrel, Sanjay Bhandari, Ramesh Rana, Nikesh Shrestha, Ram Bahadur Gurung, Cristian Linte, Angus Watson, Yash Raj Shrestha, Binod Bhattarai

MICCAI 2025

See full paper: arXiv

Vision-Language Models (VLMs), designed to interpret medical images and generate clinical text, can sometimes produce descriptions that do not match the visual content, known as hallucinations. To address this in gastrointestinal (GI) image analysis, the researchers created Gut‑VLM, a dataset built in two stages: initially generating reports using ChatGPT for Kvasir‑v2 images (which may contain hallucinations), followed by expert review to correct and tag inaccuracies. Rather than solely fine‑tuning VLMs to generate descriptive reports, they propose hallucination‑aware fine‑tuning, training models to detect and correct hallucinations. This approach outperformed traditional report‑generation fine‑tuning, and the work establishes a new benchmark for evaluating VLM fidelity in GI image analysis.

From out-of-distribution detection to hallucination-aware vision-language models and federated learning under real-world constraints, NAAMII continues to work on practical clinical challenges in medical AI.

Four MMLL Papers Accepted at MICCAI & MIUA 2025