Selected work

Representative Papers

Tokenizer Transplantation: Mitigating Autoregressive Collapse in Edge-Efficient Bengali ASR

Jul 10, 2026

This work addresses the challenges of word fragmentation and autoregressive collapse in lightweight autoregressive speech recognition models when applied to morphologically rich, non-Latin scripts such as Bengali, primarily caused by English-centric byte-level tokenizers. To overcome this without requiring full model re-pretraining, the authors propose a cross-script lexical transplantation method: replacing the decoder’s byte-level vocabulary with a Bengali-specific WordPiece vocabulary derived from BanglaBERT and rescaling the embedding matrix to better align with the target language’s linguistic structure. This approach enables the first efficient and reproducible tokenizer replacement for compact ASR architectures like Moonshine. Evaluated on the Lipi-Ghor dataset, the method achieves a word error rate of 21.54%, a real-time factor of 0.0053, an 85.8% reduction in autoregressive sequence length, and a substantial drop in word fertility from 9.16 to 1.30.

0 citationsRead paper

Spatial Neighboring Scattering Transform: A Cross-Channel Amplitude Coupling Measure for EEG Connectivity

Jul 09, 2026

Existing EEG connectivity methods predominantly rely on phase synchronization, rendering them susceptible to volume conduction artifacts and overlooking amplitude coupling information. This work proposes the Spatially Neighboring Scattering Transform (SNST), which extends wavelet scattering to multichannel settings to jointly characterize inter-channel amplitude envelope coupling and its cross-frequency modulation structure. For the first time, this approach systematically reveals EEG amplitude coupling and demonstrates its periodic regulation by slow rhythms. Applied to the BCI Competition IV-2a dataset, SNST consistently identifies significant coupling in the centro-parietal region. The resulting connectivity patterns exhibit no overlap with those derived from PLI or wPLI, thereby validating the distinctiveness of amplitude coupling and offering a novel perspective on brain connectivity that complements traditional phase-synchronization-based analyses.

0 citationsRead paper

Wavelet Scattering Transform for Interpretable Schizophrenia Biomarker Discovery and Classification from Resting-State EEG

Jul 06, 2026

Existing resting-state EEG biomarkers for schizophrenia lack objectivity and clinical translatability, often neglecting amplitude modulation dynamics and cross-frequency coupling while suffering from temporal data leakage. This study proposes a leakage-free, interpretable analytical framework that, for the first time, employs higher-order wavelet scattering transform (WST) to extract multiscale features capturing both amplitude modulation and cross-frequency coupling. Integrating leave-one-subject-out (LOSO) cross-validation, SHAP-based interpretability, and multiple comparison correction, the approach achieves a classification accuracy of 90.48% (AUC = 0.9339, sensitivity = 95.56%) under strict subject-independent evaluation. The analysis identifies a novel EEG biomarker centered on the P3 electrode and enriched in gamma-band activity, offering enhanced clinical relevance and neurophysiological interpretability.

0 citationsRead paper

Bi-cLSTM: Residual-Corrected Bidirectional LSTM for Aero-Engine RUL Estimation

Feb 28, 2026

This study addresses the challenge of insufficient accuracy and robustness in remaining useful life (RUL) prediction for aircraft engines under varying operating conditions and noisy measurements. To this end, the authors propose a bidirectional residual-corrected LSTM (Bi-cLSTM) model that integrates bidirectional temporal modeling with an adaptive residual correction mechanism. The approach further incorporates a condition-aware preprocessing pipeline—comprising operating-condition segmentation-based normalization, feature selection, and exponential smoothing—to significantly enhance generalization and noise resilience across complex, multi-regime scenarios. Experimental results on all four NASA C-MAPSS datasets demonstrate that the proposed method outperforms existing LSTM-based baselines, achieving state-of-the-art performance, with particularly pronounced gains in multi-condition settings.

0 citationsRead paper

CASR-Net: An Image Processing-focused Deep Learning-based Coronary Artery Segmentation and Refinement Network for X-ray Coronary Angiogram

Oct 31, 2025

This study addresses low segmentation accuracy, fragmented small vascular branches, and high false-positive rates in X-ray coronary angiography images. We propose a three-stage deep learning framework: (1) multi-channel preprocessing integrating CLAHE with an improved Ben Graham method to enhance contrast and suppress noise; (2) a backbone segmentation network leveraging a DenseNet121 encoder and a Self-ONN decoder to strengthen feature representation; and (3) a contour refinement module embedded to improve vascular boundary continuity and topological consistency. Evaluated via five-fold cross-validation on two public datasets, our method achieves IoU = 61.43%, Dice Similarity Coefficient (DSC) = 76.10%, and clDice = 79.36%, outperforming state-of-the-art models. The framework delivers robust, clinically relevant vessel segmentation, thereby supporting early diagnosis and precise treatment planning for coronary artery disease.

0 citationsRead paper

Recent publications

Latest Papers

Tokenizer Transplantation: Mitigating Autoregressive Collapse in Edge-Efficient Bengali ASR

Jul 10, 2026

This work addresses the challenges of word fragmentation and autoregressive collapse in lightweight autoregressive speech recognition models when applied to morphologically rich, non-Latin scripts such as Bengali, primarily caused by English-centric byte-level tokenizers. To overcome this without requiring full model re-pretraining, the authors propose a cross-script lexical transplantation method: replacing the decoder’s byte-level vocabulary with a Bengali-specific WordPiece vocabulary derived from BanglaBERT and rescaling the embedding matrix to better align with the target language’s linguistic structure. This approach enables the first efficient and reproducible tokenizer replacement for compact ASR architectures like Moonshine. Evaluated on the Lipi-Ghor dataset, the method achieves a word error rate of 21.54%, a real-time factor of 0.0053, an 85.8% reduction in autoregressive sequence length, and a substantial drop in word fertility from 9.16 to 1.30.

0 citationsRead paper

Spatial Neighboring Scattering Transform: A Cross-Channel Amplitude Coupling Measure for EEG Connectivity

Jul 09, 2026

Existing EEG connectivity methods predominantly rely on phase synchronization, rendering them susceptible to volume conduction artifacts and overlooking amplitude coupling information. This work proposes the Spatially Neighboring Scattering Transform (SNST), which extends wavelet scattering to multichannel settings to jointly characterize inter-channel amplitude envelope coupling and its cross-frequency modulation structure. For the first time, this approach systematically reveals EEG amplitude coupling and demonstrates its periodic regulation by slow rhythms. Applied to the BCI Competition IV-2a dataset, SNST consistently identifies significant coupling in the centro-parietal region. The resulting connectivity patterns exhibit no overlap with those derived from PLI or wPLI, thereby validating the distinctiveness of amplitude coupling and offering a novel perspective on brain connectivity that complements traditional phase-synchronization-based analyses.

0 citationsRead paper

Wavelet Scattering Transform for Interpretable Schizophrenia Biomarker Discovery and Classification from Resting-State EEG

Jul 06, 2026

Existing resting-state EEG biomarkers for schizophrenia lack objectivity and clinical translatability, often neglecting amplitude modulation dynamics and cross-frequency coupling while suffering from temporal data leakage. This study proposes a leakage-free, interpretable analytical framework that, for the first time, employs higher-order wavelet scattering transform (WST) to extract multiscale features capturing both amplitude modulation and cross-frequency coupling. Integrating leave-one-subject-out (LOSO) cross-validation, SHAP-based interpretability, and multiple comparison correction, the approach achieves a classification accuracy of 90.48% (AUC = 0.9339, sensitivity = 95.56%) under strict subject-independent evaluation. The analysis identifies a novel EEG biomarker centered on the P3 electrode and enriched in gamma-band activity, offering enhanced clinical relevance and neurophysiological interpretability.

0 citationsRead paper

Bi-cLSTM: Residual-Corrected Bidirectional LSTM for Aero-Engine RUL Estimation

Feb 28, 2026

This study addresses the challenge of insufficient accuracy and robustness in remaining useful life (RUL) prediction for aircraft engines under varying operating conditions and noisy measurements. To this end, the authors propose a bidirectional residual-corrected LSTM (Bi-cLSTM) model that integrates bidirectional temporal modeling with an adaptive residual correction mechanism. The approach further incorporates a condition-aware preprocessing pipeline—comprising operating-condition segmentation-based normalization, feature selection, and exponential smoothing—to significantly enhance generalization and noise resilience across complex, multi-regime scenarios. Experimental results on all four NASA C-MAPSS datasets demonstrate that the proposed method outperforms existing LSTM-based baselines, achieving state-of-the-art performance, with particularly pronounced gains in multi-condition settings.

0 citationsRead paper

CASR-Net: An Image Processing-focused Deep Learning-based Coronary Artery Segmentation and Refinement Network for X-ray Coronary Angiogram

Oct 31, 2025

This study addresses low segmentation accuracy, fragmented small vascular branches, and high false-positive rates in X-ray coronary angiography images. We propose a three-stage deep learning framework: (1) multi-channel preprocessing integrating CLAHE with an improved Ben Graham method to enhance contrast and suppress noise; (2) a backbone segmentation network leveraging a DenseNet121 encoder and a Self-ONN decoder to strengthen feature representation; and (3) a contour refinement module embedded to improve vascular boundary continuity and topological consistency. Evaluated via five-fold cross-validation on two public datasets, our method achieves IoU = 61.43%, Dice Similarity Coefficient (DSC) = 76.10%, and clDice = 79.36%, outperforming state-of-the-art models. The framework delivers robust, clinically relevant vessel segmentation, thereby supporting early diagnosis and precise treatment planning for coronary artery disease.

0 citationsRead paper