Multidimensional Convolutional Neural Networks (CNNs) and Large Language Models (LLMs) for Classification of Varicose Veins and Associated Chronic Diseases
Authors: Darshan P R, Roopa D M, Thriveni S K, Anitha A B
Varicose veins (VV) and chronic venous diseases (CVD) affect a significant portion of the global population, often leading to pain, edema, and in advanced cases, ulceration. Accurate classification of disease severity is essential for timely intervention and optimal treatment planning. This study proposes a novel multimodal framework that integrates multidimensional Convolutional Neural Networks (CNNs) and Large Language Models (LLMs) for the automated classification of varicose veins and related chronic conditions. The CNN module is designed to process high-resolution vascular imaging data, including Doppler ultrasound and venography scans, while the LLM component analyzes unstructured clinical text such as physician notes and electronic health records. The fusion of visual and textual features enables robust classification based on the CEAP (Clinical-Etiological-Anatomical-Pathophysiological) system and ICD-10 codes. Experimental results demonstrate that the proposed hybrid model outperforms unimodal approaches in terms of accuracy, sensitivity, and clinical relevance. This research highlights the potential of deep learning-based multimodal systems in enhancing diagnostic precision and decision support in vascular medicine.
Introduction
Chronic venous diseases (CVD), including varicose veins (VV), are common vascular disorders affecting quality of life and potentially leading to serious complications. Diagnosis and staging often rely on subjective clinical assessments and imaging, with the CEAP classification system as a standard but challenging to apply consistently.
Traditional machine learning methods have been used for vein imaging analysis, but they struggle with generalization. Recent advances in deep learning, especially Convolutional Neural Networks (CNNs) for imaging and Large Language Models (LLMs) for clinical text, offer promising tools for improving diagnosis. However, few studies combine these modalities for vascular disease classification.
This study proposes a novel hybrid framework that integrates multidimensional CNNs analyzing Doppler ultrasound and MR/CT images with LLMs extracting clinical information from patient records. Using a late-fusion approach, the model jointly leverages imaging and text data to classify CVD according to CEAP stages and ICD-10 codes.
Trained on a multimodal dataset of 1,200 patient cases annotated by specialists, the model outperformed single-modality baselines in accuracy (91.5%), F1-score (0.90), and ROC-AUC (0.94), especially for mid-stage disease classification. The fusion approach effectively addresses limitations of imaging or text alone by combining complementary information.
The research contributes (1) a multidimensional CNN tailored for vascular imaging, (2) fine-tuning of a biomedical LLM for clinical text, and (3) a multimodal fusion strategy that enhances classification accuracy and robustness. This approach aims to aid clinicians in making more consistent and accurate diagnoses, improving patient outcomes and setting a foundation for future multimodal AI applications in vascular and chronic disease management.
Conclusion
This study presented a novel multimodal deep learning approach for the classification of varicose veins and chronic venous diseases using multidimensional Convolutional Neural Networks (CNNs) and a Large Language Model (LLM). By integrating volumetric imaging data with unstructured clinical narratives, the model achieved enhanced accuracy and interpretability in assigning CEAP classifications—a critical step in diagnosis and treatment planning. The proposed system demonstrated superior performance compared to unimodal baselines, achieving a classification accuracy of 91.5%, F1-score of 0.90, and ROC-AUC of 0.94. These results highlight the complementary nature of imaging and textual data in understanding complex vascular conditions. While CNNs effectively captured anatomical and structural features from 3D imaging, the LLM processed nuanced symptom descriptions and clinical history, enabling a more holistic representation of patient condition.
Moreover, the attention mechanisms and visualization tools employed in the model enhanced clinical interpretability, allowing practitioners to better understand the rationale behind each classification. This transparency is vital for gaining trust in AI-assisted decision-making tools in healthcare settings. Despite certain limitations, such as data imbalance in advanced CEAP stages and variations in imaging protocols, the study establishes a strong foundation for future work in automated vascular diagnostics. Further expansion of the dataset, incorporation of longitudinal data, and real-time clinical validation could elevate this model into a powerful tool for vascular disease screening and monitoring. In conclusion, the integration of multidimensional CNNs with LLMs offers a promising pathway toward accurate, interpretable, and scalable classification of venous disorders in clinical practice.
References
[1] Varicose Vein Detection and Real-Time Integration using Faster R-CNN Algorithm by T Sriranjani; K. P Senthil Kumar; G Thiruksha published in IEEE
[2] Varicose Veins Chronic Venous Diseases Image Classification Using Multidimensional Convolutional Neural Networks Published in 2022 6th International Conference on Devices, Circuits and Systems (ICDCS) IEEE M. Roshni Thanka; E. Bijolin Edwin; R. Priscilla Joy; S. Jeba Priya; V. Ebenezer
[3] Medical Image Analysis and Classification for Varicose Veins ByJyoti Yogesh Deshmukh, Vijay U Rathod, Yogesh Kisan Mali, Rachna Sable Edition1st Edition First Published 2024
[4] VARICOEASE – A Soothebelt for Varicose veins Publisher: IEEE Nanditha Krishna; Surabhi A S; Saritha Pal; Aishwarya Prabhu; Sahana P Link:
[5] Automatic Segmentation and Detection System for Varicocele in Supine Position Publisher: IEEE Omar Alzoubi; Mohammad Abu Awad; Ayman M. Abdalla
[6] IQGAP1 promotes the phenotypic switch of vascular smooth muscle by myocardin pathway: a potential target for varicose vein by Xianchen Huang 1, Yiqi Jin 1, Dayong Zhou 1, Guoxiong Xu 1, Jian Huang 1, Liming Shen 1
[7] A Comprehensive Review on Varicose Veins and its Implications using AI Methods. M. J., Prasanna Kumar; Prakash H. N.
[8] Detection of Chronic Venous Disease Publisher: IEEE Rutika Katkar; Akshata Desai; Arati Patil; Nikita Lengare; Komal Jadhav; Mahesh Chavan