Selected work

Representative Papers

RadLite: Multi-Task LoRA Fine-Tuning of Small Language Models for CPU-Deployable Radiology AI

May 01, 2026

This work addresses the challenge of deploying large language models in resource-constrained clinical settings due to their high computational demands. The authors propose RadLite, a framework that leverages LoRA-based fine-tuning of compact models—Qwen2.5-3B and Qwen3-4B—combined with multi-task training and GGUF quantization to enable efficient inference across nine radiology tasks. This study demonstrates for the first time that lightly fine-tuned small models can effectively perform multi-task radiological analysis and operate entirely on CPU with modest memory requirements (1.8–2.4 GB) and throughput of 4–8 tokens per second. Compared to zero-shot baselines, RadLite achieves substantial improvements: +53% in RADS classification accuracy, +60% in natural language inference (NLI), and +89% in N staging. Furthermore, ensembling Qwen2.5 and Qwen3 yields complementary strengths and state-of-the-art performance in structured generation and information extraction tasks.

0 citationsRead paper

CLIP Architecture for Abdominal CT Image-Text Alignment and Zero-Shot Learning: Investigating Batch Composition and Data Scaling

Apr 15, 2026

This work investigates the impact of training batch composition on vision–language representation learning in 3D medical imaging, specifically focusing on aligning abdominal CT scans with radiology reports. Leveraging the Merlin dual-encoder architecture and symmetric InfoNCE loss, the study systematically analyzes how the ratio of normal to abnormal samples and dataset scale influence zero-shot performance. Notably, it reveals— for the first time in 3D medical imaging—that implicit diversity from random sampling consistently outperforms explicit class-balancing strategies, challenging conventional assumptions about data balancing. Experimental results demonstrate that the proposed approach achieves a zero-shot macro F1 score of 74.45% across 30 pathological findings; explicit balancing uniformly degrades performance by 2.4–2.8 points, while increasing data scale yields only sublinear gains.

0 citationsRead paper

TrackletGPT: A Language-like GPT Framework for White Matter Tract Segmentation

Jan 20, 2026

White matter tract segmentation is crucial for brain connectomics and clinical applications, yet existing methods suffer from limited generalizability due to inter-subject variability and differences in acquisition protocols. This work proposes TrackletGPT—the first end-to-end framework to adapt a GPT-like architecture for tract segmentation—by decomposing streamlines into tracklet sequences and leveraging self-attention mechanisms together with 3D structural priors to effectively model both the commonalities and variabilities of fiber tracts. Evaluated on the TractoInferno and Human Connectome Project (HCP) datasets, TrackletGPT consistently outperforms current state-of-the-art methods, achieving superior cross-dataset performance in terms of DICE, Overlap, and Overreach metrics. This study provides the first demonstration of the generalizability and efficacy of language-model-inspired paradigms in neuroimaging segmentation tasks.

0 citationsRead paper

TractRLFusion: A GPT-Based Multi-Critic Policy Fusion Framework for Fiber Tractography

Jan 20, 2026

This study addresses the persistent challenge in tractography of accurately reconstructing white matter pathways while suppressing spurious connections. To this end, the authors propose a novel multi-strategy fusion framework based on deep reinforcement learning, marking the first application of a GPT-inspired architecture to this domain. The approach integrates a two-stage training data selection scheme with a multi-critic fine-tuning mechanism to enhance both reconstruction accuracy and anatomical plausibility. Comprehensive experiments on the HCP, ISMRM, and TractoInferno datasets demonstrate that the proposed method significantly outperforms established classical techniques, single-strategy reinforcement learning baselines, and state-of-the-art deep reinforcement learning approaches in terms of both quantitative metrics and neuroanatomical fidelity.

0 citationsRead paper

Performance of a Deep Learning-Based Segmentation Model for Pancreatic Tumors on Public Endoscopic Ultrasound Datasets

Jan 09, 2026arXiv.org

This study addresses the challenge of subjective variability in endoscopic ultrasound (EUS) image segmentation for pancreatic cancer, where automated and objective tumor boundary delineation remains lacking. The authors propose a novel approach that, for the first time, employs a Vision Transformer as the backbone network within the USFM segmentation framework to perform automatic segmentation of pancreatic tumors on grayscale EUS images standardized to 512×51 pixels. Evaluated via five-fold cross-validation and external testing, the model achieves Dice scores of 0.651 and 0.657, respectively, with specificity exceeding 97%, demonstrating strong generalizability and robustness across datasets. The analysis also highlights persistent challenges in multi-region missegmentation under data heterogeneity, offering critical insights for future methodological refinements.

0 citationsRead paper

Recent publications

Latest Papers

RadLite: Multi-Task LoRA Fine-Tuning of Small Language Models for CPU-Deployable Radiology AI

May 01, 2026

This work addresses the challenge of deploying large language models in resource-constrained clinical settings due to their high computational demands. The authors propose RadLite, a framework that leverages LoRA-based fine-tuning of compact models—Qwen2.5-3B and Qwen3-4B—combined with multi-task training and GGUF quantization to enable efficient inference across nine radiology tasks. This study demonstrates for the first time that lightly fine-tuned small models can effectively perform multi-task radiological analysis and operate entirely on CPU with modest memory requirements (1.8–2.4 GB) and throughput of 4–8 tokens per second. Compared to zero-shot baselines, RadLite achieves substantial improvements: +53% in RADS classification accuracy, +60% in natural language inference (NLI), and +89% in N staging. Furthermore, ensembling Qwen2.5 and Qwen3 yields complementary strengths and state-of-the-art performance in structured generation and information extraction tasks.

0 citationsRead paper

CLIP Architecture for Abdominal CT Image-Text Alignment and Zero-Shot Learning: Investigating Batch Composition and Data Scaling

Apr 15, 2026

This work investigates the impact of training batch composition on vision–language representation learning in 3D medical imaging, specifically focusing on aligning abdominal CT scans with radiology reports. Leveraging the Merlin dual-encoder architecture and symmetric InfoNCE loss, the study systematically analyzes how the ratio of normal to abnormal samples and dataset scale influence zero-shot performance. Notably, it reveals— for the first time in 3D medical imaging—that implicit diversity from random sampling consistently outperforms explicit class-balancing strategies, challenging conventional assumptions about data balancing. Experimental results demonstrate that the proposed approach achieves a zero-shot macro F1 score of 74.45% across 30 pathological findings; explicit balancing uniformly degrades performance by 2.4–2.8 points, while increasing data scale yields only sublinear gains.

0 citationsRead paper

TrackletGPT: A Language-like GPT Framework for White Matter Tract Segmentation

Jan 20, 2026

White matter tract segmentation is crucial for brain connectomics and clinical applications, yet existing methods suffer from limited generalizability due to inter-subject variability and differences in acquisition protocols. This work proposes TrackletGPT—the first end-to-end framework to adapt a GPT-like architecture for tract segmentation—by decomposing streamlines into tracklet sequences and leveraging self-attention mechanisms together with 3D structural priors to effectively model both the commonalities and variabilities of fiber tracts. Evaluated on the TractoInferno and Human Connectome Project (HCP) datasets, TrackletGPT consistently outperforms current state-of-the-art methods, achieving superior cross-dataset performance in terms of DICE, Overlap, and Overreach metrics. This study provides the first demonstration of the generalizability and efficacy of language-model-inspired paradigms in neuroimaging segmentation tasks.

0 citationsRead paper

TractRLFusion: A GPT-Based Multi-Critic Policy Fusion Framework for Fiber Tractography

Jan 20, 2026

This study addresses the persistent challenge in tractography of accurately reconstructing white matter pathways while suppressing spurious connections. To this end, the authors propose a novel multi-strategy fusion framework based on deep reinforcement learning, marking the first application of a GPT-inspired architecture to this domain. The approach integrates a two-stage training data selection scheme with a multi-critic fine-tuning mechanism to enhance both reconstruction accuracy and anatomical plausibility. Comprehensive experiments on the HCP, ISMRM, and TractoInferno datasets demonstrate that the proposed method significantly outperforms established classical techniques, single-strategy reinforcement learning baselines, and state-of-the-art deep reinforcement learning approaches in terms of both quantitative metrics and neuroanatomical fidelity.

0 citationsRead paper

Performance of a Deep Learning-Based Segmentation Model for Pancreatic Tumors on Public Endoscopic Ultrasound Datasets

Jan 09, 2026arXiv.org

This study addresses the challenge of subjective variability in endoscopic ultrasound (EUS) image segmentation for pancreatic cancer, where automated and objective tumor boundary delineation remains lacking. The authors propose a novel approach that, for the first time, employs a Vision Transformer as the backbone network within the USFM segmentation framework to perform automatic segmentation of pancreatic tumors on grayscale EUS images standardized to 512×51 pixels. Evaluated via five-fold cross-validation and external testing, the model achieves Dice scores of 0.651 and 0.657, respectively, with specificity exceeding 97%, demonstrating strong generalizability and robustness across datasets. The analysis also highlights persistent challenges in multi-region missegmentation under data heterogeneity, offering critical insights for future methodological refinements.

0 citationsRead paper