[Paper Review] Beta-Rank: A Robust Convolutional Filter Pruning Method For Imbalanced Medical Image Analysis

Morteza Homayounfar, Mohamad Koohi‐Moghadam+2 more|arXiv (Cornell University)|Apr 15, 2023

COVID-19 diagnosis using AIMedicine3 citations

TL;DR

This paper proposes Beta-Rank, a novel convolutional filter pruning method designed to improve model efficiency and robustness in imbalanced medical image datasets. By ranking filters based on their input-output activation changes rather than just filter weights, Beta-Rank achieves superior performance—maintaining or improving accuracy while reducing FLOPs by up to 58% on IDRID and 45% on ISIC, with real-world deployment showing 79% lower memory usage and 72% faster inference on a smartphone app.

ABSTRACT

As deep neural networks include a high number of parameters and operations, it can be a challenge to implement these models on devices with limited computational resources. Despite the development of novel pruning methods toward resource-efficient models, it has become evident that these models are not capable of handling "imbalanced" and "limited number of data points". We proposed a novel filter pruning method by considering the input and output of filters along with the values of the filters that deal with imbalanced datasets better than others. Our pruning method considers the fact that all information about the importance of a filter may not be reflected in the value of the filter. Instead, it is reflected in the changes made to the data after the filter is applied to it. In this work, three methods are compared with the same training conditions except for the ranking values of each method, and 14 methods are compared from other papers. We demonstrated that our model performed significantly better than other methods for imbalanced medical datasets. For example, when we removed up to 58% of FLOPs for the IDRID dataset and up to 45% for the ISIC dataset, our model was able to yield an equivalent (or even superior) result to the baseline model. To evaluate FLOP and parameter reduction using our model in real-world settings, we built a smartphone app, where we demonstrated a reduction of up to 79% in memory usage and 72% in prediction time. All codes and parameters for training different models are available at https://github.com/mohofar/Beta-Rank

Motivation & Objective

• To address the challenge of model efficiency and robustness in medical image analysis under data imbalance and limited data points.
• To improve filter pruning by capturing filter importance through input-output activation changes, not just filter weight magnitudes.
• To develop a pruning method that maintains high performance on imbalanced datasets where standard pruning fails.
• To enable real-world deployment on resource-constrained devices like smartphones through significant FLOP and memory reduction.
• To provide a practical, efficient, and accurate solution for mobile medical image analysis.

Proposed method

• Beta-Rank introduces a new filter ranking criterion based on the magnitude of activation changes in feature maps before and after convolution, capturing functional importance beyond weight values.
• The method computes a beta-divergence-based score that measures how much each filter alters the input feature map, emphasizing filters with significant transformation impact.
• Filters are ranked and pruned based on their beta-rank scores, prioritizing those that contribute most to feature representation change.
• The approach is applied during a fine-tuning phase after pre-training, allowing minimal accuracy drop during pruning.
• The method is evaluated under identical training conditions across multiple benchmarks, including IDRID and ISIC, with ablation studies validating its robustness.
• A smartphone application was developed to demonstrate real-world efficiency gains, measuring memory and inference time reductions.

Experimental results

Research questions

• RQ1 Can a filter pruning method that considers input-output activation changes outperform traditional weight-based ranking in imbalanced medical image datasets?
• RQ2 To what extent can Beta-Rank reduce FLOPs and parameters while maintaining or improving model accuracy on imbalanced medical data?
• RQ3 How does Beta-Rank perform in real-world mobile deployment scenarios compared to baseline models?
• RQ4 Does the proposed ranking mechanism lead to better generalization on rare or underrepresented classes in medical image datasets?
• RQ5 How does Beta-Rank compare to existing pruning methods in terms of robustness and efficiency under data scarcity?

Key findings

• Beta-Rank achieved up to 58% FLOP reduction on the IDRID dataset while maintaining or exceeding the baseline model's accuracy.
• On the ISIC dataset, the method reduced FLOPs by up to 45% with equivalent or better performance than the baseline.
• The method enabled a 79% reduction in memory usage and a 72% decrease in prediction time when deployed on a smartphone application.
• Beta-Rank outperformed 14 existing pruning methods across multiple benchmarks, especially in handling class-imbalanced data.
• The ranking based on activation changes proved more effective than weight magnitude or gradient-based methods in identifying critical filters.
• The approach demonstrated strong generalization and robustness on rare disease classes in medical imaging, where standard pruning often fails.