[Paper Review] Evolution-based Fine-tuning of CNNs for Prostate Cancer Detection

Khashayar Namdar, Isha Gujrathi+2 more|arXiv (Cornell University)|Nov 4, 2019

Advanced Neural Network Applications20 references4 citations

TL;DR

This paper proposes an evolution-based fine-tuning method using a Genetic Algorithm (GA) to optimize the fully connected (FC) layers of a pre-trained CNN for prostate cancer detection on 6-channel diffusion-weighted MRI. By applying GA at the layer level instead of individual weights, the method improves test AUC by 9.3% (from 0.707 to 0.773) on 1,334 test slices, achieving a significant performance gain without retraining the feature extraction layers.

ABSTRACT

Convolutional Neural Networks (CNNs) have been used for automated detection of prostate cancer where Area Under Receiver Operating Characteristic (ROC) curve (AUC) is usually used as the performance metric. Given that AUC is not differentiable, common practice is to train the CNN using a loss functions based on other performance metrics such as cross entropy and monitoring AUC to select the best model. In this work, we propose to fine-tune a trained CNN for prostate cancer detection using a Genetic Algorithm to achieve a higher AUC. Our dataset contained 6-channel Diffusion-Weighted MRI slices of prostate. On a cohort of 2,955 training, 1,417 validation, and 1,334 test slices, we reached test AUC of 0.773; a 9.3% improvement compared to the base CNN model.

Motivation & Objective

• To address the suboptimal AUC performance in CNN-based prostate cancer detection despite using standard backpropagation with differentiable loss functions.
• To overcome the limitation of SGD in optimizing non-differentiable metrics like AUC by introducing evolutionary computation to the classifier head.
• To reduce computational cost of GA-based CNN optimization by operating at the layer level rather than individual weights.
• To demonstrate that evolutionary fine-tuning of FC layers can significantly improve AUC without retraining convolutional feature extractors.
• To validate the method on a real-world DW-MRI dataset of 5,706 prostate slices from 414 patients.

Proposed method

• A pre-trained CNN with 3 convolutional and 3 fully connected (FC) layers was used, with weights initialized via Xavier and uniform distributions.
• The CNN was first trained using cross-entropy loss and SGD with a fixed learning rate, momentum, and L2 penalty.
• A Genetic Algorithm was applied exclusively to the FC layers, treating each layer as a single entity for crossover and mutation operations.
• The initial GA population consisted of 512 classifiers: one from the trained CNN and 511 randomly initialized ones.
• Fitness was evaluated based on AUC on the training set, with selection, crossover, and mutation (1% probability) applied over generations.
• Optimization was halted when validation and training AUC stopped improving, with final model selected based on highest validation AUC.

Experimental results

Research questions

• RQ1 Can a genetic algorithm effectively optimize the fully connected layers of a CNN to improve AUC in prostate cancer detection from DW-MRI?
• RQ2 Does layer-level GA mutation and crossover reduce computational cost while maintaining performance gains compared to node-level optimization?
• RQ3 Can evolutionary fine-tuning outperform standard SGD training in terms of AUC, even when the feature extraction layers are frozen?
• RQ4 Is the improvement in AUC statistically significant and robust across training, validation, and independent test sets?
• RQ5 How does the proposed method compare to conventional training with AUC monitoring in terms of final model performance?

Key findings

• The proposed GA-based fine-tuning improved test AUC from 0.707 (SGD-trained model) to 0.773, representing a 9.3% relative improvement.
• The best model was achieved in just three generations of the GA, with validation AUC peaking at 0.815 and test AUC at 0.773.
• The method achieved optimal performance in only 10 minutes on a GeForce GTX 1060 GPU, demonstrating computational feasibility.
• The GA improved AUC on the training set from 0.867 (SGD) to 0.877, indicating better generalization.
• The approach successfully optimized the non-differentiable AUC metric by bypassing backpropagation constraints through evolutionary search.
• The layer-level GA approach reduced computational overhead compared to node-level optimization, making it practical for low-end hardware.