Selected work

Representative Papers

Lesion Segmentation in FDG-PET/CT Using Swin Transformer U-Net 3D: A Robust Deep Learning Framework

Jan 06, 2026

This study addresses the limited accuracy of automatic lesion segmentation in FDG-PET/CT imaging by proposing SwinUNet3D, a novel framework that effectively integrates the shifted-window self-attention mechanism of Swin Transformer with the skip-connection architecture of 3D U-Net. This integration enables simultaneous modeling of global contextual information and preservation of fine anatomical details, while also optimizing multimodal PET/CT feature fusion. The method substantially enhances detection of small and irregular lesions and reduces false-positive rates. Evaluated on the AutoPET III dataset, SwinUNet3D achieves a Dice coefficient of 0.88 and an IoU of 0.78—significantly outperforming standard 3D U-Net (Dice 0.48, IoU 0.32)—and demonstrates faster inference speed.

1 citationsRead paper

Towards Explainability of SLMs by investigating Token Level Activation

May 21, 2026

Existing attention-based interpretability methods often overemphasize structurally salient but semantically weak tokens—such as punctuation—failing to reveal the true semantic mechanisms of language models. To address this, this work proposes Activation Flow Network (AFN), a lightweight, model-agnostic interpretability framework that leverages the L2 norm of hidden states from the 8th layer of BERT as a measure of semantic salience. Tokens are partitioned into high- and low-activation groups using an empirical interquartile threshold. Experimental results demonstrate that semantically rich tokens consistently occupy the high-activation group and dominate representational evolution, confirming that the 8th layer serves as a critical region where structural and semantic information converge. This approach significantly enhances model transparency and semantic focus.

0 citationsRead paper

A New Technique for AI Explainability using Feature Association Map

May 12, 2026

Current AI systems struggle to gain trust in high-stakes scenarios due to their lack of transparency, necessitating more accurate explanation methods. This work proposes FAMeX, an algorithm that introduces Feature Association Maps (FAMs) to model inter-feature dependencies using graph theory, thereby challenging the common assumption of feature independence in existing explainable AI (XAI) approaches. By capturing the contextual importance of features in classification tasks more faithfully, FAMeX provides a more realistic representation of feature contributions. Experimental results across eight benchmark datasets demonstrate that FAMeX significantly outperforms mainstream XAI methods such as Permutation Feature Importance (PFI) and SHAP in evaluating feature importance, exhibiting superior explanatory power and practical utility.

0 citationsRead paper

CLAReSNet: When Convolution Meets Latent Attention for Hyperspectral Image Classification

Nov 15, 2025

Hyperspectral image classification faces challenges including high-dimensional spectral features, complex spectral-spatial correlations, and small-sample class imbalance. To address these, we propose a hybrid convolutional–latent-attention network. Our method introduces an adaptive latent bottleneck mechanism and multi-scale spectral latent attention, reducing computational complexity to $O(T log T cdot D)$. Additionally, we design a spectral-domain encoding layer that integrates multi-scale convolutional residual blocks, bidirectional RNNs, and an enhanced convolutional attention module to enable hierarchical cross-modal feature fusion. Evaluated on the Indian Pines and Salinas datasets, our model achieves overall accuracies of 99.71% and 99.96%, respectively—substantially outperforming state-of-the-art methods including HybridSN, SSRN, and SpectralFormer. These results demonstrate the proposed model’s superior accuracy, computational efficiency, and robustness under limited training samples and imbalanced class distributions.

0 citationsRead paper

Improving Predictive Confidence in Medical Imaging via Online Label Smoothing

Oct 22, 2025

Deep learning models for medical image classification often exhibit overconfident predictions, undermining clinical reliability. To address this, we propose Online Label Smoothing (OLS), a label regularization method that dynamically models the distribution of prediction confidence and adaptively attenuates soft-label weights for non-target classes, thereby modulating supervision strength without requiring a teacher model or additional parameters. OLS enhances feature embedding compactness and separability across diverse architectures—including ResNet-50, MobileNetV2, and VGG-19. On the RadImageNet benchmark, OLS outperforms hard labels, standard label smoothing, and teacher-free knowledge distillation baselines in Top-1 and Top-5 accuracy. Moreover, it significantly improves model calibration, reducing Expected Calibration Error (ECE) by up to 38%. Thus, OLS simultaneously advances discriminative performance and predictive trustworthiness—critical for safe clinical deployment.

0 citationsRead paper

Recent publications

Latest Papers

Towards Explainability of SLMs by investigating Token Level Activation

May 21, 2026

Existing attention-based interpretability methods often overemphasize structurally salient but semantically weak tokens—such as punctuation—failing to reveal the true semantic mechanisms of language models. To address this, this work proposes Activation Flow Network (AFN), a lightweight, model-agnostic interpretability framework that leverages the L2 norm of hidden states from the 8th layer of BERT as a measure of semantic salience. Tokens are partitioned into high- and low-activation groups using an empirical interquartile threshold. Experimental results demonstrate that semantically rich tokens consistently occupy the high-activation group and dominate representational evolution, confirming that the 8th layer serves as a critical region where structural and semantic information converge. This approach significantly enhances model transparency and semantic focus.

0 citationsRead paper

A New Technique for AI Explainability using Feature Association Map

May 12, 2026

Current AI systems struggle to gain trust in high-stakes scenarios due to their lack of transparency, necessitating more accurate explanation methods. This work proposes FAMeX, an algorithm that introduces Feature Association Maps (FAMs) to model inter-feature dependencies using graph theory, thereby challenging the common assumption of feature independence in existing explainable AI (XAI) approaches. By capturing the contextual importance of features in classification tasks more faithfully, FAMeX provides a more realistic representation of feature contributions. Experimental results across eight benchmark datasets demonstrate that FAMeX significantly outperforms mainstream XAI methods such as Permutation Feature Importance (PFI) and SHAP in evaluating feature importance, exhibiting superior explanatory power and practical utility.

0 citationsRead paper

Lesion Segmentation in FDG-PET/CT Using Swin Transformer U-Net 3D: A Robust Deep Learning Framework

Jan 06, 2026

This study addresses the limited accuracy of automatic lesion segmentation in FDG-PET/CT imaging by proposing SwinUNet3D, a novel framework that effectively integrates the shifted-window self-attention mechanism of Swin Transformer with the skip-connection architecture of 3D U-Net. This integration enables simultaneous modeling of global contextual information and preservation of fine anatomical details, while also optimizing multimodal PET/CT feature fusion. The method substantially enhances detection of small and irregular lesions and reduces false-positive rates. Evaluated on the AutoPET III dataset, SwinUNet3D achieves a Dice coefficient of 0.88 and an IoU of 0.78—significantly outperforming standard 3D U-Net (Dice 0.48, IoU 0.32)—and demonstrates faster inference speed.

1 citationsRead paper

CLAReSNet: When Convolution Meets Latent Attention for Hyperspectral Image Classification

Nov 15, 2025

Hyperspectral image classification faces challenges including high-dimensional spectral features, complex spectral-spatial correlations, and small-sample class imbalance. To address these, we propose a hybrid convolutional–latent-attention network. Our method introduces an adaptive latent bottleneck mechanism and multi-scale spectral latent attention, reducing computational complexity to $O(T log T cdot D)$. Additionally, we design a spectral-domain encoding layer that integrates multi-scale convolutional residual blocks, bidirectional RNNs, and an enhanced convolutional attention module to enable hierarchical cross-modal feature fusion. Evaluated on the Indian Pines and Salinas datasets, our model achieves overall accuracies of 99.71% and 99.96%, respectively—substantially outperforming state-of-the-art methods including HybridSN, SSRN, and SpectralFormer. These results demonstrate the proposed model’s superior accuracy, computational efficiency, and robustness under limited training samples and imbalanced class distributions.

0 citationsRead paper

Improving Predictive Confidence in Medical Imaging via Online Label Smoothing

Oct 22, 2025

Deep learning models for medical image classification often exhibit overconfident predictions, undermining clinical reliability. To address this, we propose Online Label Smoothing (OLS), a label regularization method that dynamically models the distribution of prediction confidence and adaptively attenuates soft-label weights for non-target classes, thereby modulating supervision strength without requiring a teacher model or additional parameters. OLS enhances feature embedding compactness and separability across diverse architectures—including ResNet-50, MobileNetV2, and VGG-19. On the RadImageNet benchmark, OLS outperforms hard labels, standard label smoothing, and teacher-free knowledge distillation baselines in Top-1 and Top-5 accuracy. Moreover, it significantly improves model calibration, reducing Expected Calibration Error (ECE) by up to 38%. Thus, OLS simultaneously advances discriminative performance and predictive trustworthiness—critical for safe clinical deployment.

0 citationsRead paper