[Paper Review] Plexus Convolutional Neural Network (PlexusNet): A novel neural network architecture for histologic image analysis

Okyaz Eminağa, Mahmoud Abbas+6 more|arXiv (Cornell University)|Aug 23, 2019

AI in cancer detection4 citations

TL;DR

PlexusNet is a novel, lightweight convolutional neural network architecture designed for histologic image analysis that achieves state-of-the-art classification performance with significantly reduced parameter count. It outperforms larger models in calibration and clinical utility, delivering AUCs of 0.963 for prostate cancer and 0.978 for breast cancer metastasis detection while being up to 200× smaller than ResNet18.

ABSTRACT

Different convolutional neural network (CNN) models have been tested for their application in histological image analyses. However, these models are prone to overfitting due to their large parameter capacity, requiring more data or valuable computational resources for model training. Given these limitations, we introduced a novel architecture (termed PlexusNet). We utilized 310 Hematoxylin and Eosin stained (H&E) annotated histological images of prostate cancer cases from TCGA-PRAD and Stanford University and 398 H&E whole slides images from the Camelyon 2016 challenge. PlexusNet-architecture -derived models were compared to models derived from several existing "state of the art" architectures. We measured discrimination accuracy, calibration, and clinical utility. An ablation study was conducted to study the effect of each component of PlexusNet on model performance. A well-fitted PlexusNet-based model delivered comparable classification performance (AUC: 0.963) in distinguishing prostate cancer from healthy tissues, although it was at least 23 times smaller, had a better model calibration and clinical utility than the comparison models. A separate smaller PlexusNet model accurately detected slides with breast cancer metastases (AUC: 0.978); it helped reduce the slide number to examine by 43.8% without consequences, although its parameter capacity was 200 times smaller than ResNet18. We found that the partitioning of the development set influences the model calibration for all models. However, with PlexusNet architecture, we could achieve comparable well-calibrated models trained on different partitions. In conclusion, PlexusNet represents a novel model architecture for histological image analysis that achieves classification performance comparable to other models while providing orders-of-magnitude parameter reduction.

Motivation & Objective

• To address overfitting and high computational demands in histologic image analysis by designing a smaller, more efficient neural network architecture.
• To improve model calibration and clinical utility compared to existing state-of-the-art models.
• To evaluate the impact of data partitioning on model calibration and ensure robustness across different training splits.
• To achieve high classification accuracy with minimal parameter capacity in whole-slide image analysis.
• To enable efficient screening by reducing the number of slides requiring manual review without loss of diagnostic accuracy.

Proposed method

• Proposed a novel neural network architecture, PlexusNet, featuring a hierarchical, multi-scale feature aggregation mechanism to enhance representational capacity with fewer parameters.
• Employed a partitioned training strategy to study the influence of data split on model calibration and generalization.
• Designed a compact variant of PlexusNet for low-resource settings, optimized for inference efficiency and parameter efficiency.
• Trained and evaluated models on two datasets: 310 H&E-stained prostate cancer slides from TCGA-PRAD and Stanford, and 398 whole slides from the Camelyon 2016 challenge.
• Conducted ablation studies to isolate the contribution of each architectural component to overall performance.
• Used AUC, calibration metrics, and clinical utility scores to compare PlexusNet against established architectures like ResNet18.

Experimental results

Research questions

• RQ1 Can a smaller neural network architecture achieve comparable classification performance to larger state-of-the-art models in histologic image analysis?
• RQ2 How does the partitioning of the training set affect model calibration across different architectures?
• RQ3 To what extent does the PlexusNet architecture improve model calibration and clinical utility compared to existing models?
• RQ4 Can a compact version of PlexusNet effectively reduce the number of slides requiring manual review in cancer screening?
• RQ5 What is the contribution of each architectural component in PlexusNet to its overall performance and efficiency?

Key findings

• A well-fitted PlexusNet-based model achieved an AUC of 0.963 in distinguishing prostate cancer from healthy tissue, matching the performance of larger models.
• The same model was at least 23 times smaller than comparison models and demonstrated better calibration and clinical utility.
• A smaller PlexusNet variant detected breast cancer metastases with an AUC of 0.978 and reduced the number of slides to examine by 43.8% without loss of accuracy.
• This compact model had a parameter count 200 times smaller than ResNet18 while maintaining high performance.
• PlexusNet achieved well-calibrated models across different data partitioning schemes, unlike baseline models that were sensitive to split variations.
• The ablation study confirmed that each architectural component contributed meaningfully to performance, especially in feature aggregation and parameter efficiency.