[Paper Review] C-Net: A Reliable Convolutional Neural Network for Biomedical Image Classification

Hosein Barzekar, Zeyun Yu|arXiv (Cornell University)|Oct 30, 2020

AI in cancer detection51 references4 citations

TL;DR

This paper proposes C-Net, a novel deep convolutional neural network that concatenates multiple networks—Outer, Middle, and Inner—to improve biomedical image classification. By using six feature-extracting networks in the Outer and Middle stages to feed into the Inner classifier, C-Net achieves 99.34% average accuracy and 99.52% F1-score on the BreakHis dataset and 100% precision and zero misclassifications on the Osteosarcoma dataset, outperforming prior models across all metrics.

ABSTRACT

Cancers are the leading cause of death in many countries. Early diagnosis plays a crucial role in having proper treatment for this debilitating disease. The automated classification of the type of cancer is a challenging task since pathologists must examine a huge number of histopathological images to detect infinitesimal abnormalities. In this study, we propose a novel convolutional neural network (CNN) architecture composed of a Concatenation of multiple Networks, called C-Net, to classify biomedical images. The model incorporates multiple CNNs including Outer, Middle, and Inner. The first two parts of the architecture contain six networks that serve as feature extractors to feed into the Inner network to classify the images in terms of malignancy and benignancy. The C-Net is applied for histopathological image classification on two public datasets, including BreakHis and Osteosarcoma. To evaluate the performance, the model is tested using several evaluation metrics for its reliability. The C-Net model outperforms all other models on the individual metrics for both datasets and achieves zero misclassification. This approach has the potential to be extended to additional classification tasks, as experimental results demonstrate utilizing extensive evaluation metrics.

Motivation & Objective

• To address the challenge of accurate and reliable automated classification of histopathological images in cancer diagnosis.
• To overcome limitations in existing CNN architectures that struggle with image heterogeneity, staining variability, and feature extraction in low-contrast or complex tissue structures.
• To develop a robust deep learning model that generalizes well across diverse datasets, magnification levels, and disease types.
• To evaluate the model using a comprehensive set of metrics beyond accuracy, including F1-score, MCC, precision, recall, and specificity.
• To demonstrate the potential of a concatenated multi-network architecture for improving feature extraction and classification reliability in medical imaging.

Proposed method

• C-Net employs a hierarchical architecture composed of three main components: Outer, Middle, and Inner networks, where Outer and Middle networks serve as feature extractors.
• The Outer and Middle networks each consist of six parallel Networks-in-Network (NIN) modules, enabling multi-scale and multi-level feature extraction from input histopathological images.
• Features from the Outer and Middle networks are concatenated and fed into the Inner network, which performs the final binary classification between benign and malignant tumors.
• The model uses standard deep learning components such as convolutional layers, batch normalization, ReLU activation, and global average pooling, with cross-entropy loss for training.
• The architecture is trained end-to-end on two public datasets: BreakHis (7,909 breast cancer images at 40X–400X magnification) and Osteosarcoma (1,144 high-resolution 1024×1024 images).
• Evaluation is performed using multiple metrics including accuracy, F1-score, Matthews Correlation Coefficient (MCC), precision, recall, and specificity to ensure robustness and reliability.

Experimental results

Research questions

• RQ1 Can a concatenated multi-network CNN architecture improve feature representation and classification performance in biomedical image analysis compared to standard deep CNNs?
• RQ2 How does C-Net perform across different magnification levels on the BreakHis dataset, particularly in terms of recall and F1-score?
• RQ3 To what extent does C-Net reduce misclassification rates compared to state-of-the-art models on both BreakHis and Osteosarcoma datasets?
• RQ4 Does the use of multiple parallel feature extractors (Outer and Middle networks) lead to higher robustness and generalization in diverse histopathological imaging conditions?
• RQ5 Can the proposed architecture achieve near-perfect performance on challenging biomedical classification tasks, such as detecting viable tumors in osteosarcoma images?

Key findings

• C-Net achieved 99.34% average accuracy on the BreakHis dataset, which is 1.71% higher than the highest prior average accuracy reported in the literature.
• On BreakHis, the model attained a 99.52% average F1-score, surpassing the previous best average F1-score of 97.67% by 1.85 percentage points.
• The model achieved 100% recall on 40X and 400X magnification levels of the BreakHis dataset, indicating no missed malignant cases.
• On the Osteosarcoma dataset, C-Net achieved 100% precision in two of three classification categories (NCT vs VT and VT vs NT), with 98.25% average precision across all categories.
• The model recorded 99.18% specificity on the Osteosarcoma dataset, which is 5.35% higher than the best prior result reported in the literature.
• C-Net achieved 99.23% MCC on 200X images from BreakHis, the highest among all models evaluated, and only 8 total misclassifications across all magnification levels, demonstrating exceptional reliability.