TLDR: This research introduces a comprehensive multi-task deep learning framework for bladder cancer diagnosis from cystoscopic images. It includes a classification model (EfficientNet-B0 with CBAM) achieving 93.28% accuracy, a segmentation model (ResNet34-UNet++) with a 0.9091 Dice coefficient for precise lesion delineation, and an exploratory model (ConvNeXt-Tiny) for molecular subtyping (HER-2, Ki-67, p53). The framework is integrated into a user-friendly, Gradio-based online platform offering real-time, AI-assisted diagnostic support, multi-format image uploads, and bilingual interfaces, aiming to improve the accuracy, efficiency, and accessibility of bladder cancer diagnostics.
Bladder cancer is a significant health concern globally, being one of the most common cancers of the urinary tract. Current diagnostic methods, primarily clinical cystoscopy, rely heavily on the expertise of physicians, which can lead to variations and subjectivity in diagnosis. This highlights a critical need for more objective, accurate, and efficient computational tools to improve how bladder cancer is detected and managed.
A recent study introduces an innovative multi-task deep learning framework and an accompanying online platform designed to enhance bladder cancer diagnosis using cystoscopic images. This integrated approach aims to provide more reliable, efficient, and accessible diagnostic support for clinicians.
The core of this research is a sophisticated deep learning framework that tackles three key diagnostic areas:
Tumor Classification
The framework includes a robust classification model built on EfficientNet-B0, enhanced with a Convolutional Block Attention Module (CBAM). This model is trained to distinguish between tumor and non-tumor regions in cystoscopic images. It achieved impressive results, with an accuracy of 93.28%, an F1-score of 82.05%, and an AUC (Area Under the Curve) of 96.41% on internal validation. Even when tested on external datasets from different hospitals and equipment, it maintained strong performance, demonstrating its reliability across various clinical settings.
Lesion Segmentation
To precisely identify and outline tumor boundaries, the study developed an advanced segmentation model. This model is based on a ResNet34-UNet++ architecture, incorporating self-attention mechanisms and attention gating. It achieved an exceptional Dice coefficient of 0.9091, indicating high accuracy in delineating lesion boundaries. This precision is crucial for surgical planning and ensuring complete tumor removal.
Also Read:
- Evaluating Virtual DES Images for Breast Lesion Classification in Mammography
- AI Models Enhance Knee MRI Analysis for Injury Detection
Molecular Subtyping
Beyond visual identification, the research also explores the feasibility of predicting molecular markers like HER-2 and Ki-67 directly from endoscopic images using a ConvNeXt-Tiny model. While still an exploratory task with performance constrained by dataset size, permutation testing confirmed that the model could identify statistically significant patterns linked to these molecular expressions. This opens a promising path for non-invasive, real-time biomarker assessment, which could lead to more personalized treatment strategies.
To make these powerful AI models accessible to clinicians, the researchers developed a Gradio-based online diagnostic platform. This user-friendly platform allows for multi-format image uploads, offers bilingual interfaces (Chinese and English), and enables dynamic adjustment of diagnostic thresholds. It integrates all the developed models, providing real-time feedback on tumor classification, segmentation overlays, and optional molecular marker predictions. The platform is designed to seamlessly fit into clinical workflows, acting as a decision support tool for experienced urologists and an educational resource for junior physicians.
Pilot tests with urologists showed high satisfaction with the platform’s clarity, speed, and the interpretability of its results, particularly the Grad-CAM visualizations that highlight regions influencing the model’s decisions. The average inference time per image is under 500 milliseconds, ensuring near-instantaneous feedback.
This multi-task framework and integrated online tool represent a significant advancement in intelligent bladder cancer diagnosis. By improving clinical reliability, supporting early tumor detection, and enabling real-time diagnostic feedback, these contributions mark a substantial step toward AI-assisted decision-making in urology. For more details, you can refer to the full research paper here.
