A Calibrated Multiexit Neural Network for Detecting Urothelial Cancer Cells.

Area of Science:

• Medical Imaging
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Deep convolutional networks (CNNs) are effective for medical imaging diagnostics.
• Cytopathology, particularly urinary cytopathology, is an underexplored area for deep learning applications compared to histology.

Purpose of the Study:

• To propose a novel deep learning model for cancer detection in urinary cytopathology screening images.
• To address the challenge of small-scale datasets common in real-world cytopathology scenarios.
• To improve model calibration, ensuring confidence levels align with true probabilities.

Main Methods:

• Development of a novel deep learning model using multi-output neural networks for efficient training on small datasets.
• Implementation of techniques to enhance model calibration, including focal loss and temperature scaling.
• Evaluation on a novel urinary cytopathology dataset.

Main Results:

• The proposed model demonstrates significantly improved accuracy and calibration compared to a baseline deep convolutional network.
• The combination of focal loss, multiple outputs, and temperature scaling proved effective.
• The model is capable of efficient training on small-scale datasets.

Conclusions:

• The novel deep learning approach offers a promising solution for cancer detection in urinary cytopathology.
• Improved calibration is crucial for reliable diagnostic tools in medical imaging.
• The model's performance on small datasets highlights its potential for real-world clinical application.