Artificial Intelligence-Based Classification and Segmentation of Bladder Cancer in Cystoscope Images.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Cystoscopy is crucial for bladder cancer diagnosis but struggles with ambiguous lesions like carcinoma in situ (CIS).
• Limitations in identifying subtle lesions contribute to high bladder cancer recurrence rates.
• Deep learning shows promise for enhancing medical image analysis in urology.

Purpose of the Study:

• To apply VGG19 and Deeplab v3+ deep learning models for classifying and segmenting cystoscope images of bladder cancer.
• To evaluate the performance of these models in identifying cancerous lesions and their morphologies.

Main Methods:

• Utilized a dataset of 772 patients' cystoscope images, annotated by experienced urologists.
• Developed a VGG19 model for lesion classification based on morphology and pathology.
• Employed a Deeplab v3+ model for segmenting different bladder cancer morphologies.
• Used sparse categorical cross-entropy and dice coefficient loss functions for classification and segmentation, respectively.

Main Results:

• The VGG19 classification model achieved an accuracy of 0.912.
• The Deeplab v3+ segmentation model attained an Intersection over Union (IoU) of 0.833 and binary accuracy of 0.951.
• Visual inspection confirmed high concordance between Deeplab v3+ segmentations and expert annotations.

Conclusions:

• VGG19 and Deeplab v3+ demonstrated robust performance in classifying and segmenting bladder cancer from cystoscopic images.
• These deep learning models offer valuable tools for bladder cancer research.
• The models show potential to assist clinicians in diagnosing bladder cancer more effectively.