Singapore

Recent advancements in deep learning have enabled expert-level performance in medical imaging for disease classification, but their black-box decision making processes limit trust in them and their wide-spread clinical deployment. While Explainable Artificial Intelligence (XAI) methods aim to bridge this gap, studies focus on 2D data or pre-processed research datasets that overlook the role of medical imaging pre-processing operations which is an essential component of real-world 3D medical imaging workflows. To address this limitation, we propose XIMED, a systematic and predictive model–centered framework for evaluating explainability under realistic medical pre-processing conditions for volumetric medical data. The framework integrates five volumetric pre-processing variants and ten post-hoc attribution methods, evaluated through three complementary criteria: Correctness, Contrastivity, and Completeness, which together evaluate explanation dependence on model input, internal structure, and output behavior. Across more than 300 experimental configurations, XIMED reveals that gradient-based methods, such as Integrated Gradients and Blur Integrated Gradients, provide the most consistent and model-aligned explanations, while noise-based approaches like SmoothGrad and VarGrad are less sensitive to model behavior. These findings emphasize the importance of clinically realistic evaluation pipelines for reliable explainability in 3D medical imaging.

AAAI 2026

XIME3D : A Systematic Framework for Evaluating Explainable AI in 3D Medical Imaging under CT Image Pre-processing Variations

medical imaging preprocessing

xai evaluation

3d medical imaging

explainable ai

# Welcome everyone to ACL 2022!

The 60th Annual Meeting of the Association for Computational Linguistics is taking place May 22-27, 2022 as a hybrid event, in Dublin and online. We are happy to welcome all of you to this anniversary edition with an almost 50-50 in-person and virtual participation. 
The main conference program features oral presentations, in-person and virtual posters and demo sessions, a plenary session for our best paper presentations and awards, three amazing keynote events and two new initiatives of invited talks: Spotlight Talks for Young Rising Stars (STIRS) and The Next Big Idea Talks. Posters (including Findings of ACL 2022) and demos are grouped by areas for both the in-person and the virtual sessions. For the virtual component, the talks will be on Zoom and the posters and the demos will be in GatherTown. The Student Research Workshop will have an oral session and a poster session as part of Poster Session 1. The program also features eight Tutorials and 28 Workshops. 

 
We wish you a wonderful conference! 
[**The ACL 2022 Organizing Committee**](https://www.2022.aclweb.org/organisers)
 
[**Conference Handbook**](https://drive.google.com/file/d/1_BUCMfhMVrjG9E2e71aHdHeE28KSje0l/view?usp=sharing) 
[**Mini Handbook**](https://drive.google.com/file/d/1qlBKl0wzmlVF1oCeMQl3BahLd9nLP5Ce/view?usp=sharing) 
[**Posters and Demo guides**](https://drive.google.com/file/d/1UucMAoCNncIOaH1rMMDa0owuG9qgvJTG/view?usp=sharing)

ACL 2022

The Association for Computational Linguistics (ACL) is the premier international scientific and professional society for people working on computational problems involving human language, a field often referred to as either computational linguistics or natural language processing (NLP). 

technical paper

We are pleased to announce the Fortieth AAAI Conference on Artificial Intelligence (AAAI-26), which will be held in Singapore EXPO from January 20 to January 27, 2026.

The purpose of the AAAI conference series is to promote research in Artificial Intelligence (AI) and foster scientific exchange between researchers, practitioners, scientists, students, and engineers across the entirety of AI and its affiliated disciplines. AAAI-26 will feature technical paper presentations, special tracks, invited speakers, workshops, tutorials, poster sessions, senior member presentations, competitions, and exhibit programs, and a range of other activities to be announced.

To access this event page, you need to log in with the **email address you registered with**. Access credentials will be sent to your email from Underline - subject line "Welcome to AAAI 2026". Please be sure to check your spam email folder if you do not see an email confirmation right away.

Please log in

To access this event page, you are required to register.
Please complete your registration to continue.

We recommend reading [**the registration information**](https://aaai.org/conference/aaai/aaai-26/registration/) first.

**Online Registration Form**: https://aaai.getregistered.net/conference-2026 

Registration Required

We are pleased to announce the Fortieth AAAI Conference on Artificial Intelligence (AAAI-26), which will be held in Singapore EXPO from January 20 to January 27, 2026.

Aortic stenosis (AS) remains a major cardiovascular challenge, where early ECG indicators often elude conventional diagnostic approaches. Despite the success of deep learning in AS detection, clinical translation is impeded by limited interpretability of model decisions. This study investigates how neural network architecture influences explainability and interpretability in AS classification. Three models—ResNet-18 (CNN-based), LSTM, and ViT-MAE—were trained on an open-access cardiomechanical signals dataset of 100 valvular heart disease patients. Using Integrated Gradients, all models achieved strong performance (ResNet-18: 97.23%, LSTM: 98.96%, ViT-MAE: 88.56%) but displayed distinct attention mechanisms: ResNet-18 utilized global ECG features, LSTM captured temporal QRS dependencies, and ViT-MAE emphasized T-wave and QT interval features. All models consistently attended to clinically meaningful regions aligned with known AS physiology. These results reveal that explainability is architecture-dependent and advocate for architectureaware interpretation strategies to foster transparency, reliability, and clinical trust in cardiovascular AI applications.

Architecture-Aware Explainability in ECG Analysis: A Case Study of Aortic Stenosis Detection with ResNet18, LSTM and ViT-MAE ECG

We study meal-level inference of the carbohydrate calorie ratio (CCR), the fraction of total calories attributable to net carbohydrates, from early postprandial glucose responses (PPGR) recorded by continuous glucose monitors (CGMs). Using the date-shifted CGMACROS v1.0.0 dataset, we introduce GlucoGrapher, a deployment-minded pipeline that (i) aligns PPGR around meal onset, (ii) performs sensor-aware fusion of Dexcom and Libre traces with explicit discordance detection and gating, (iii) applies a dual preprocessing view of the signal (absolute mg/dL and percent change relative to baseline) with per-subject PPGR standardization to mitigate inter-individual scale shifts, and (iv) combines heterogeneous, and sometimes missing—out-of-fold predictions using fold-wise, mask-aware nonnegative meta-learning with isotonic calibration to avoid leakage. We augment PPGR shape descriptors (peak, time-to-peak, IAUC windows, slopes, late-ratio features) with lightweight behavior (steps, heart rate) and compact subject context (BMI, HbA1c strata, selected fasting labs, and up to eight microbiome principal components). Evaluated with 5-fold GroupKFold by participant over meals, GlucoGrapher attains RMSE , , , and Pearson . Performance is consistent across HbA1c-defined strata (Healthy/PreDM/T2D), indicating robustness to baseline glycemic status. Ablations show that the mask-aware meta-ensemble provides a substantive lift over calibrated tree baselines, highlighting the value of reliability-aware sensor fusion, dual preprocessing, and leak-free calibration. Framing CCR as a bounded, interpretable target in enables actionable CGM-only feedback without perfect food logging, supporting retrospective coaching and prospective planning.

GlucoGrapher: Sensor-Aware CGM Fusion with Mask-Aware Meta-Ensembles for Predicting Carbohydrate Caloric Ratio from Postprandial Glucose

Radiology worklists are typically processed on a first-in, first-out (FIFO) basis, even when studies differ greatly in clinical urgency. We propose a pragmatic alternative: using calibrated probabilities of intracranial hemorrhage (ICH) to prioritize head CT exams for earlier reading. Using the public RSNA-ICH dataset, we train slice-level detectors, aggregate them to the exam level, apply post-hoc calibration, and feed these scores into a transparent discrete-event simulator of the reading queue. The simulator quantifies how triage benefit reduction in median time-to-read (TTR) for ICH; scales with classifier AUC, workload (arrival rate), staffing, prevalence, and calibration. Across realistic loads, score-based prioritization yields substantial TTR reductions for ICH with minimal delay to non-ICH studies. We release a configuration-driven, reproducible pipeline that translates AI risk scores into operational metrics (minutes saved), enabling safe and data-driven evaluation before PACS/RIS deployment.

Translating Classifier Scores into Clinical Impact: Calibrated Risk and Queueing Simulation for AI-Assisted Radiology Worklist Triage

Multimodal medical image fusion aims to integrate complementary information from different imaging modalities to enhance clinical diagnosis and surgical navigation. While deep learning-based approaches have significantly advanced fusion quality over traditional methods by leveraging powerful feature extraction, challenges such as blurring, noise, and artifacts persist in the fused results. To address these issues, we propose a novel fusion framework that incorporates an entropy-based attention module to emphasize salient image regions. Our architecture is designed in a multi-scale manner, utilizing an adaptive gating mechanism to effectively extract and combine salient features across different scales. Additionally, we introduce a Top-K token vision transformer to enable efficient global feature extraction while reducing computational overhead by restricting the context space. We further demonstrate the effectiveness of our fused representations in the downstream task of oocyte quality prediction, showing improved accuracy over individual focal images as well as over other approaches. Extensive experiments on diverse medical imaging datasets demonstrate that our method achieves competitive performance compared to state-of-the-art techniques, both quantitatively and visually. Ablation studies underscore the importance of each proposed component.

TAFIE: Transformer-Assisted Fusion with Integrated Entropy Attention for Multimodal Medical Imaging

Chest X-rays are a widely-used, cost-effective imaging modality for medical investigations; however, they encode distinctive biometric signatures that enable identification attacks. We introduce \emph{PrivDiff-Net}, a novel diffusion-based framework that addresses critical privacy vulnerabilities in chest X-rays through anatomical biometric features while preserving essential diagnostic utility for clinical applications. Our approach introduces two modules in a latent diffusion framework: (1) a Selective Attribute Suppression (SAS) module that removes sensitive identity cues using orthogonal projection in cross-attention, and (2) a Selective Privacy Guidance (SPG) loss that discourages identity features while preserving diagnostic information during diffusion. Quantitative results show that PrivDiff-Net achieves near-random identification (AUC: 47%) while maintaining high diagnostic accuracy (AUC: 78%). It effectively suppresses sensitive attributes and produces high-quality anonymized CXRs, validated by clinicians for diagnostic utility. These results establish \emph{PrivDiff-Net} as a new benchmark for privacy-preserving Chest X-rays, providing a practical solution for secure data sharing in collaborative research environments while enabling ethical deployment of AI systems in healthcare where transparency and patient privacy are critical.

Hide Identity, Preserve Pathology: Diffusion-Based Anonymization for Chest X-rays

Conversational AI is starting to support real clinical work, but most evaluation methods miss how compliance depends on the full course of a conversation. We introduce Obligatory-Information Phase Structured Compliance Evaluation (OIP–SCE), an evaluation method that checks whether every required clinical obligation is met, in the right order, with clear evidence for clinicians to review. This makes complex rules practical and auditable, helping close the gap between technical progress and what healthcare actually needs. We demonstrate the method in two case studies (respiratory history, benefits verification) and show how phase-level evidence turns policy into shared, actionable steps. By giving clinicians control over what to check and engineers a clear specification to implement, OIP–SCE provides a single, auditable evaluation surface that aligns AI capability with clinical workflow and supports routine, safe use.

All Required, In Order: Phase-Level Evaluation for AI–Human Dialogue in Healthcare and Beyond

Semi-supervised learning (SSL) has become a promising direction for medical image segmentation, enabling models to learn from limited labeled data alongside abundant unlabeled samples. However, existing SSL approaches for multi-modal medical imaging often struggle to exploit the complementary information between modalities due to semantic discrepancies and misalignment across MRI sequences. To address this, we propose a novel semi-supervised multi-modal framework that explicitly enhances modality-specific representations and facilitates adaptive cross-modal information fusion. Specifically, we introduce a Modality-specific Enhancing Module (MEM) to strengthen semantic cues unique to each modality via channel-wise attention, and a learnable Complementary Information Fusion (CIF) module to adaptively exchange complementary knowledge between modalities. The overall framework is optimized using a hybrid objective combining supervised segmentation loss and cross-modal consistency regularization on unlabeled data. Extensive experiments on the BraTS 2019 (HGG subset) demonstrate that our method consistently outperforms strong semi-supervised and multi-modal baselines under 1%, 5%, and 10% labeled data settings, achieving significant improvements in both Dice and Sensitivity scores. Ablation studies further confirm the complementary effects of our proposed MEM and CIF in bridging cross-modality discrepancies and improving segmentation robustness under scarce supervision.

Modality-Specific Enhancement and Complementary Fusion for Semi-Supervised Multi-Modal Brain Tumor Segmentation

Recent advances in foundation models have enabled their integration into high-stakes clinical settings, particularly in computational pathology, where domain-specialized FMs demonstrate strong generalization. However, real-world deployment is constrained by their poorly calibrated uncertainty awareness and degraded performance in low-data regimes requiring few-shot adaptation strategies, leading to unreliable and inefficient diagnostic workflows. Conformal Prediction (CP) is an uncertainty quantification framework that offers distribution‑free, finite‑sample coverage guarantees for ensuring safer deployment in such settings. In this work, we explore the integration of various CP methods with pathology foundation models using three few-shot adaption strategies for classification tasks across two datasets. To assess the clinical effectiveness of these approaches, we propose four novel metrics aimed at improving clinical reliability and alleviating diagnostic workload in few-shot settings. Our results demonstrate that Conformal Prediction methods enhance the reliability of pathology foundation models and offer actionable uncertainty estimates to enable safe and efficient deployment in few-shot pathological classification workflows, with the LAC method achieving the best overall performance.

Towards Reliable Few-Shot Adaptation of Pathology Foundation Models via Conformal Prediction

Recent advancements in Large Language Models (LLMs) and Large Vision Language Models (LVLMs) have enabled general-purpose systems to demonstrate promising capabilities in complex reasoning tasks, including those in the medical domain. Medical Visual Question Answering (MedVQA) has particularly benefited from these developments. However, despite Bangla being one of the most widely spoken languages globally, there exists no established MedVQA benchmark for it. To address this gap, we introduce BanglaMedVQA, a dataset comprising clinically validated image–question–answer pairs, along with a comprehensive evaluation of current foundation models on this resource. Consistent with prior findings that report low performance of current models on English MedVQA benchmarks, our analysis reveals that Bangla performance is substantially lower, reflecting the challenges inherent to low-resource languages. Even top-performing models such as Gemini and GPT-4.1 mini fail to accurately answer specialized diagnostic questions, indicating severe limitations in fine-grained medical reasoning. Although certain open-source models, such as Gemma-3, occasionally outperform these models in general categories, they too struggle with clinically complex questions underscoring the urgent need for top-notch evaluation method.

How Good LLMs Are at Answering Bangla Medical Visual Questions? Dataset and Benchmarking

We study an ultrasound-first, radiation-preserving policy for developmental dysplasia of the hip (DDH) that requests a radiograph only when needed. We (i) pretrain modality-specific encoders (ResNet‑18) with SimSiam on a large unlabelled registry (37{,}186 ultrasound; 19{,}546 radiographs), (ii) freeze the backbones and fit small, measurement-faithful heads on DDH relevant landmarks and measurements (iii) calibrate a one-sided conformal deferral rule on ultrasound predictions that guarantees coverage on a held-out calibration set. Ultrasound heads predict Graf and femoral head coverage; X-ray heads predict acetabular index (AI), center–edge (CE) angle, and IHDI grade . On our held out labeled evaluation set, ultrasound measurement error is modest (e.g., MAE 9.7, coverage MAE 14.0%), while radiographic probes achieve AI and CE MAEs of 7.6 and 8.9, respectively. The calibrated US-only policy is explored across rule families (alpha-only; alpha OR coverage; alpha AND coverage), uncertainty inflation factors, and per-utility trade-offs using decision-curve analysis. Conservative settings yield high coverage (e.g., 0.90 for ) with near-zero US-only rates; permissive settings (e.g., alpha OR coverage at larger deltas) achieve non-zero US-only throughput with expected coverage trade-offs. The result is a simple, reproducible pipeline that turns limited labels into interpretable measurements and tunable selective imaging curves suitable for clinical handoff and future external validation.

Downloads

Next from AAAI 2026

Architecture-Aware Explainability in ECG Analysis: A Case Study of Aortic Stenosis Detection with ResNet18, LSTM and ViT-MAE ECG

Similar lecture

A Novel Framework Based on Medical Concept Driven Attention for Explainable Medical Code Prediction via External Knowledge

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from AAAI 2026

Architecture-Aware Explainability in ECG Analysis: A Case Study of Aortic Stenosis Detection with ResNet18, LSTM and ViT-MAE ECG

Similar lecture

A Novel Framework Based on Medical Concept Driven Attention for Explainable Medical Code Prediction via External Knowledge

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES

Downloads