Medico 2025 Challenge: Advancing Explainable GI VQA

• Medico 2025 Challenge is a competition advancing state-of-the-art Visual Question Answering and explainable AI in gastrointestinal endoscopy by requiring both accurate responses and clear, clinical explanations.
• It features dual subtasks where systems must predict clinically plausible answers from endoscopy images and generate multimodal explanations with clinical rationales and precise localization maps.
• The challenge utilizes the Kvasir-VQA-x1 dataset and strong baselines like CNN+Transformer Q-Encoder to ensure rigorous evaluation based on accuracy, BLEU, IoU, and expert review.

The Medico 2025 Challenge is a prominent competition within the MediaEval task series, targeted at advancing Visual Question Answering (VQA) and explainable artificial intelligence (XAI) for gastrointestinal (GI) endoscopy. The challenge is designed to benchmark systems that not only answer clinically meaningful questions about GI images but also generate interpretable, multimodal explanations closely aligned with medical reasoning (Gautam et al., 14 Aug 2025).

1. Challenge Structure and Objectives

Medico 2025 is structured around two interdependent subtasks:

• Visual Question Answering (VQA): Systems receive an endoscopy image and a free-text, clinically focused question and must output a discrete, clinically plausible answer.
• Multimodal Explanation Generation: Beyond answer prediction, systems are required to provide a human-interpretable explanation consisting of a clinical-language rationale and a localization map that highlights the image regions supporting the answer.

The overarching objective is to advance trustworthy, transparent medical AI by integrating robust performance with explicit, clinically coherent interpretability.

2. Problem Formulation and Technical Specification

VQA Subtask:

The central machine learning formulation posits the VQA model as a mapping $f: \mathcal{I} \times \mathcal{Q} \to \mathcal{A}$ where $\mathcal{I}$ is the set of RGB GI images, $\mathcal{Q}$ is the set of natural language questions (e.g., "What is the location of the lesion?"), and $\mathcal{A}$ is a constrained answer set (binary, multiple-choice, or short phrase). Supervised learning proceeds by minimizing the cross-entropy loss:

$$L_\mathrm{VQA} = - \sum_{i=1}^{N} \sum_{a \in \mathcal{A}} y_{i,a} \cdot \log p(a \mid I_i, Q_i)$$

with inference via maximum a posteriori selection.

Multimodal Explanation Subtask:

In tandem with answer prediction, the model must map image-question pairs $(I, Q)$ to an explanation tuple $(T, M)$ where $T$ is a clinical rationale (free text) and $M$ is a spatial mask (attention or segmentation, $M \in [0,1]^{H\times W}$) corresponding to decision-relevant regions. The extended objective is:

$\mathcal{I}$0

where $\mathcal{I}$1 is the negative log-likelihood for $\mathcal{I}$2, $\mathcal{I}$3 penalizes misalignment with expert-annotated masks, and $\mathcal{I}$4 are hyperparameters for task trade-off.

The challenge places strong emphasis on rationales referencing salient image features (e.g. "mucosal irregularity") and mask alignment with pathologic structures.

3. Data: Kvasir-VQA-x1 Dataset

The challenge is grounded in the Kvasir-VQA-x1 dataset, constructed specifically for rigorous VQA and XAI research in GI endoscopy. Key characteristics:

Split	Images	QA pairs
Training	5,600	8,400
Validation	1,400	2,100
Test	1,400	2,100

• Image Corpus: 8,400 unique RGB endoscopy images ($\mathcal{I}$5).
• Annotations: 12,600 question–answer pairs created by expert gastroenterologists.
• Question Categories:
  • Pathology Detection (e.g. “Is this ulcerative colitis?”)
  • Lesion Localization (e.g. “Where is the bleeding site?”)
  • Procedure-Related (e.g. “Which resection step is shown?”)
• Stratified Data Partitioning: Ensures balanced category and task representation across splits.

4. System Requirements and Baselines

Participants are encouraged to utilize, and the organizers provide, strong multimodal baselines:

• CNN + Transformer Q-Encoder: Combines a ResNet-152 image backbone with a BERT-style question encoder, fusing modalities via cross-modal multi-head attention.
• Concept Bottleneck Model: Predicts a dense vector of explicit, predefined clinical concepts as an intermediate layer, mapping these to both VQA outputs and textual explanations.

All models must output both an answer and an explanation for each input instance. Interpretability is enforced via explicit generation of attention maps (e.g., Grad-CAM or self-attention weights) and via concept-layer transparency.

5. Evaluation Metrics

Evaluation is conducted on both answer accuracy and the fidelity of explanations, employing both automated metrics and clinician review:

VQA Metrics:

• Accuracy: $\mathcal{I}$6
• Precision, Recall, F$\mathcal{I}$7: Per-class and macro-computed.
• BLEU$\mathcal{I}$8, METEOR: n-gram overlap/sentence alignment for rationales.
• ROUGE-L: Longest common subsequence matching for textual explanations.

Explainability Metrics:

• Spatial Alignment Score: Average Intersection-over-Union (IoU) between predicted and expert-annotated masks:

$\mathcal{I}$9

• Expert Review: Rated 1–5 (Likert scale) by clinicians for "clinical validity" and "helpfulness." The mean Likert score across test cases produces a final explainability rating.

Quantitative and qualitative assessment together ensure submitted systems uphold rigorous technical and clinical standards.

6. Participation Protocol

• Access and Registration: Data available via the Grand Challenge platform; registration and data usage agreement required.
• Code Repository: Baseline implementations and evaluation scripts are supplied (GitHub).
• Submission Format: Teams submit Docker containers comprising the required inference API (inputs: image and question; outputs: answer and explanations plus rationale mask), accompanied by a short technical report.

7. Impact and Outlook

Medico 2025 sets a new benchmark for trustworthy, explainable AI in GI endoscopy VQA. By requiring systems to unite correct prediction, spatially and semantically aligned explanations, and explicit clinical reasoning, the challenge pushes the state of the art toward practically usable, transparent clinical CDSS for endoscopic diagnosis and intervention. The integration of standardized metrics with expert review addresses both algorithmic and end-user demands for interpretability and accountability, facilitating responsible deployment of medical AI (Gautam et al., 14 Aug 2025).