Hybrid quantum image classification and federated learning for hepatic steatosis diagnosis

TL;DR

This work tackles accurate hepatic steatosis diagnosis from liver biopsy images under privacy constraints by deploying a hybrid quantum neural network (HQNN) atop a ResNet backbone and evaluating privacy-preserving federated learning (FL) across distributed hospitals. The HQNN employs a Quantum Depth-Infused (QDI) layer that processes 100 features with only 5 qubits, achieving $97\%$ accuracy, $1.8\%$ above a classical baseline, while FL maintains accuracy above $90\%$ across up to 32 clients. The study uses a 4400-image NAFLD dataset, cropped to $258\times258$, balanced across four steatosis stages, and demonstrates that HQNNs can outperform classical models on limited data with significantly fewer parameters. The combination of HQNN and FL offers a scalable, privacy-preserving framework for medical image analysis in settings with data scarcity and strict privacy requirements, with potential generalization to quantum federated learning and heatmap-based clinical tools.

Abstract

In the realm of liver transplantation, accurately determining hepatic steatosis levels is crucial. Recognizing the essential need for improved diagnostic precision, particularly for optimizing diagnosis time by swiftly handling easy-to-solve cases and allowing the expert time to focus on more complex cases, this study aims to develop cutting-edge algorithms that enhance the classification of liver biopsy images. Additionally, the challenge of maintaining data privacy arises when creating automated algorithmic solutions, as sharing patient data between hospitals is restricted, further complicating the development and validation process. This research tackles diagnostic accuracy by leveraging novel techniques from the rapidly evolving field of quantum machine learning, known for their superior generalization abilities. Concurrently, it addresses privacy concerns through the implementation of privacy-conscious collaborative machine learning with federated learning. We introduce a hybrid quantum neural network model that leverages real-world clinical data to assess non-alcoholic liver steatosis accurately. This model achieves an image classification accuracy of 97%, surpassing traditional methods by 1.8%. Moreover, by employing a federated learning approach that allows data from different clients to be shared while ensuring privacy, we maintain an accuracy rate exceeding 90%. This initiative marks a significant step towards a scalable, collaborative, efficient, and dependable computational framework that aids clinical pathologists in their daily diagnostic tasks.

Figures (3)

• Figure 1: (a) Hepatic tissue with steatosis. The image on the right shows a hepatic tissue with a severe degree of steatosis. Fat droplets are marked with black arrow. Vessels are marked with asterisk. (b) 0: liver biopsy with score 0; 1: in this field, hepatocytes have steatosis between 5 and $33\%$ (score 1); 2: liver biopsy with macrovesicular steatosis between 33 and $66\%$ with an inhomogeneous distribution of fat drops (score 2); 3: steatosis over $66\%$ (score 3). Images of [1024 $\times$ 1024] pixel. Hematoxylin and eosin, 20$\times$. (c) Architecture of a Hybrid quantum ResNet model for analyzing liver biopsy images. It starts with a classical ResNet18 neural network, followed by two fully connected layers. The information is then relayed to a QDI layer consisting of $5$ qubits and $20$ variational layers. After quantum processing, a classical 5-element vector is inputted into another fully connected layer that determines the suitability of the liver for transplantation. (d) Horizontal federated learning architecture for analyzing liver biopsy images. The cylinders represent the datasets that are used to train a classical ResNet18, both of which are physically present in the individual hospitals. The weights of the ResNet are then sent to a server (i.e., aggregation) which processes them and returns the updated weights to the individual ResNet18.
• Figure 2: (a) The graph presents the correlation between classification accuracy (depicted by green bars) and the percentage of false negative answers in classification of hepatic steatosis stage (represented by blue bars) of the classical model, relative to the class weight ratio, $\lambda$. The black bars indicate the standard deviation of values observed during model training using $5$-fold cross-validation. (b) The relationship between the classification accuracy of hepatic steatosis and the size of the training dataset for both the hybrid model (illustrated in green) and the classical model (depicted in blue) with $\lambda = 1$. The testing set contains $400$ images for all experiments depicted in this figure.
• Figure 3: Behavior of different architectures as the number of samples available in the dataset changes.