A novel self-learning framework for bladder cancer grading using histopathological images.

Area of Science:

• Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Bladder cancer incidence and mortality are rising, with muscle-invasive bladder cancer (MIBC) presenting the worst prognosis.
• Accurate grading of MIBC is crucial for effective treatment and patient outcomes.
• Current grading methods can be subjective and time-consuming.

Purpose of the Study:

• To develop and validate a self-learning framework for grading MIBC from histological images.
• To introduce a novel Deep Convolutional Embedded Attention Clustering (DCEAC) model for automated disease severity classification.
• To achieve unsupervised learning of clinically relevant patterns in MIBC histology.

Main Methods:

• Utilized a fully unsupervised, two-step learning methodology with a novel Deep Convolutional Embedded Attention Clustering (DCEAC) model.
• Applied DCEAC to high-resolution (512x512 pixel) histological images stained by immunohistochemical techniques.
• Incorporated a convolutional attention module to refine latent space features before classification.

Main Results:

• The DCEAC model achieved an average accuracy of 0.9034 in a multi-class scenario, surpassing state-of-the-art methods by 2-3%.
• Demonstrated superior performance compared to previous clustering-based methods.
• Class activation maps confirmed the model's ability to autonomously identify clinically relevant patterns without annotation.

Conclusions:

• The developed self-learning framework offers a breakthrough in MIBC grading, bridging the gap with supervised learning.
• DCEAC provides an accurate and automated method for classifying MIBC histological patterns.
• This unsupervised approach holds significant potential for improving diagnostic efficiency and accuracy in bladder cancer management.