Source-linked AI summary
Blockchain-Federated-Learning and Deep Learning Models for COVID-19 detection using CT Imaging
Rajesh Kumar, Abdullah Aman Khan, Sinmin Zhang, Jay Kumar, Ting Yang, Noorbakhash Amiri Golalirz, Zakria, Ikram Ali, Sidra Shafiq, WenYong Wang
TL;DR
COVID-19 diagnosis is challenged by testing-kit shortages, overlapping CT findings, and hospitals’ privacy constraints on data sharing. The paper proposes normalized multi-hospital CT processing with Capsule Network segmentation and classification, combined with blockchain-based federated learning. It reports improved CT-recognition performance and privacy-preserving collaborative training, while acknowledging heterogeneous data sources and a dataset collected from hospitals in Chengdu during the epidemic’s early days.
Problem
COVID-19 diagnosis and collaborative model training are constrained by testing-kit shortages, overlapping CT findings, limited confidential data access, and hospital privacy concerns.
Method
The framework normalizes heterogeneous CT data, uses SegCaps and Capsule Networks for detection, and combines hospital models through blockchain-based federated learning.
Results
The proposed framework reports better COVID-19 CT-detection performance and privacy-preserving collaboration without federated learning affecting accuracy.
Takeaways & Limitations
Hospitals can collaboratively train and share a global CT-recognition model while keeping patient data private and sharing model weights or gradients.
Takeaways & Limitations
The dataset was collected from various hospitals in Chengdu during the early days of the epidemic and required preprocessing because of data discrepancies.
Abstract
from arXiv · showhide
With the increase of COVID-19 cases worldwide, an effective way is required to diagnose COVID-19 patients. The primary problem in diagnosing COVID-19 patients is the shortage and reliability of testing kits, due to the quick spread of the virus, medical practitioners are facing difficulty identifying the positive cases. The second real-world problem is to share the data among the hospitals globally while keeping in view the privacy concerns of the organizations. Building a collaborative model and preserving privacy are major concerns for training a global deep learning model. This paper proposes a framework that collects a small amount of data from different sources (various hospitals) and trains a global deep learning model using blockchain based federated learning. Blockchain technology authenticates the data and federated learning trains the model globally while preserving the privacy of the organization. First, we propose a data normalization technique that deals with the heterogeneity of data as the data is gathered from different hospitals having different kinds of CT scanners. Secondly, we use Capsule Network-based segmentation and classification to detect COVID-19 patients. Thirdly, we design a method that can collaboratively train a global model using blockchain technology with federated learning while preserving privacy. Additionally, we collected real-life COVID-19 patients data, which is, open to the research community. The proposed framework can utilize up-to-date data which improves the recognition of computed tomography (CT) images. Finally, our results demonstrate a better performance to detect COVID-19 patients.
I. INTRODUCTION
The paper addresses COVID-19 diagnosis challenges caused by testing-kit shortages, overlapping CT findings, and limited secure access to hospital data. It proposes collaborative CT-image recognition using normalization, Capsule Networks, blockchain, and federated learning.
- I. INTRODUCTION: COVID-19 diagnosis is difficult because testing kits are scarce and CT appearances overlap with other viral and inflammatory lung diseases.CT images may show ground-glass opacity, pulmonary consolidation, rounded morphology, and peripheral lung distribution.
- I. INTRODUCTION: Hospitals face privacy and data-sharing barriers that limit collaborative training of accurate global deep-learning models.The paper identifies confidential-data availability, collaborative global-model training, and dataset collection as linked challenges.
- I. INTRODUCTION: The stated goal is accurate COVID-19 CT recognition using collaboratively trained models while preserving hospital privacy.The framework is intended to support secure sharing and global model training across hospitals.
- I. INTRODUCTION: The framework normalizes heterogeneous CT data, applies SegCaps segmentation and Capsule Network recognition, and trains a global model through blockchain-based federated learning.It is designed for data from multiple hospitals and scanner types while sharing model information rather than private patient data.
- I. INTRODUCTION: The paper introduces the CC-19 dataset, containing CT slices from 89 subjects, including 68 confirmed COVID-19 patients and 21 negative cases.The dataset contains 34,006 CT scan slices from approximately 231 CT volumes.
D. Contributions
The paper’s contributions combine heterogeneous-CT normalization, Capsule Network-based image analysis, and blockchain-enabled federated learning. It also contributes the CC-19 dataset and describes secure collaborative model sharing across hospitals.
- D. Contributions: The framework normalizes CT data from different hospitals and scanner devices before collaborative learning.The normalization process addresses heterogeneous scanner parameters and brings images to a common standard.
- D. Contributions: SegCaps segmentation and Capsule Network classification are used to detect COVID-19 patterns in lung CT scans.The workflow segments images, trains local detection models, and uses Capsule Networks for recognition.
- D. Contributions: Blockchain-based federated learning combines locally trained hospital models into a collaborative global model without sharing private patient data.Hospitals share model weights and gradients, while blockchain distributes the models over a decentralized network.
- D. Contributions: The CC-19 dataset contains 89 subjects, including 68 confirmed COVID-19 patients, and 34,006 CT scan slices.The data were collected from multiple hospitals and include approximately 231 CT scan volumes.
- D. Contributions: The proposed framework is presented as suitable for collaborative, privacy-preserving analysis of large lung-CT datasets and newer COVID-19 information.The paper describes decentralized storage and sharing as part of a real-time hospital use case.
2) Signal Normalization:
Signal normalization standardizes CT voxel intensities across hospitals using lung-window information, while the subsequent pipeline segments infection regions and classifies COVID-19 patterns with Capsule Networks.
- 2) Signal Normalization:: Different hospitals provide CT scans with heterogeneous Hounsfield Unit ranges, so lung-window parameters define voxel intensity bounds.Window level is the central signal value, and window width defines the window’s span.
- 2) Signal Normalization:: The normalized intensity maps the lung-window range to [−0.5, 0.5] to standardize the embedding space.Ioriginal denotes the input intensity and Inormalized the final intensity.
- 2) Signal Normalization:: The segmented CT images are used to train a Capsule Network for COVID-19 detection.The paper positions segmentation as the preceding stage for classification.
- 2) Signal Normalization:: The method uses 2D CT slices from XY, XZ, and YZ planes to formalize three-dimensional lung-infection views.A standardized 334×334×512 mm3 lung volume is used for segmentation.
- 2) Signal Normalization:: The segmentation formulation aggregates voxel probabilities from three-dimensional views to predict infection points.The prose defines probB as probability, B as an infection point, and g as an aggregation function for Pxy, Pyz, and Pxz.
- 2) Signal Normalization:: Capsule Networks are selected for medical-image disease detection, with dynamic routing adjusting low-level to high-level capsule distributions by agreement.The routing algorithm iteratively updates coupling-related quantities before returning capsule outputs.
C. Federated Learning to train the global model
Federated learning lets multiple hospitals collaboratively train a global model by sharing model updates rather than patient data. Random subsampling, distortion, and noise mechanisms are used to protect privacy during decentralized training.
- Federated learning combines models from multiple hospitals while keeping their patient data private.Hospitals share model weights or updates over a decentralized network rather than sharing the underlying data.
- Random subsampling selects a subset of hospitals in each communication round before distributing the global weights.The sampled hospitals calculate local updates relative to the distributed model and send those updates to the decentralized network.
- Gaussian distortion adds noise to the summed updates after sensitivity-based scaling and clipping.The update sensitivity is bounded by S, and the distorted result combines the clipped update sum with Gaussian noise.
- The distortion level is controlled because excessive noise can remove detail from the subsampled average and prevent learning improvement.The stated noise variance depends on S, σ, and m_t through S^2σ^2/m_t.
- Hospital updates are compared using variance between patients as a similarity measure.The paper defines V_c as the sum of variances in the update matrix and uses update parameters indexed by x and y.
D. Blockchain based fast and effective Federated Learning
The blockchain component stores and retrieves model-sharing transactions while keeping large, sensitive CT data at the hospitals. Privacy mechanisms, encryption, signatures, consensus, and distributed storage support collaborative training and retrieval.
- CT scans remain stored by hospitals because placing sensitive, high-volume data on a blockchain is financially and computationally expensive.The blockchain instead helps retrieve trained models and records data-sharing transactions that verify ownership.
- Each hospital records data type and size in a blockchain transaction so sharing and retrieval requests can be verified.The retrieval mechanism is intended to support collaboration among multiple hospitals without violating hospital privacy.
- Node distance and weight-matrix attributes are used to retrieve data among hospitals and generate unique hospital identifiers.The distance is defined between nodes H_i and H_j, whose attributes contribute to the retrieval calculation.
- Randomized mechanisms, including Laplace noise for local model training, are applied to protect privacy across hospital nodes.The paper describes neighboring records, sensitivity, and an outcome set for the randomized privacy mechanism.
- Consensus evaluates local models using mean absolute error before approved transactions are stored in the distributed ledger.Encrypted and signed transactions are broadcast, verified, and stored only after approval.
1) Data Sharing Process:
The data-sharing process exchanges trained model information instead of original medical images. Federated learning and blockchain are combined to support secure collaborative learning across hospitals.
- Data providers exchange learned models rather than original personal data to reduce privacy risks during sharing.The described process responds to security concerns associated with transmitting personal data to requesters.
- Blockchain and federated learning are integrated to retrieve data securely for multiple hospitals and support effective prediction.The process is presented as decentralized data sharing across hospitals worldwide.
- The proposed architecture trains a global model from locally trained hospital models while protecting the original image data.Hospitals select training data, run private federated learning, and share local model weights with the blockchain network.
A. CC-19 Dataset
The CC-19 dataset combines CT slices from multiple hospitals and scanners, with preprocessing addressing heterogeneous and inconsistent scans. It contains labeled COVID-19 cases and is publicly available for research.
- Dataset composition: 28,395 of 34,006 CT slices belong to 68 confirmed COVID-19-positive subjects, while 21 of 89 subjects were negative.The dataset is publicly available through GitHub.
- Preprocessing and heterogeneity: CT scans with discrepancies were discarded, and patients’ scans contain different numbers of slices.The text attributes slice-count variation to differences in patient height and body structure.
B. Evaluation Measures
The evaluation uses sensitivity, specificity, accuracy, and comparative deep-learning experiments to assess COVID-19 detection and segmentation. Results emphasize Capsule Network sensitivity, segmentation performance, and federated-learning behavior as provider numbers change.
- Evaluation measures: Sensitivity and specificity measure how correctly the model identifies subjects with and without disease, respectively.The paper emphasizes sensitivity because missing a COVID-19 patient can have disastrous consequences.
- Benchmark evaluation: The experiments compare VGG16, AlexNet, Inception V3, ResNet, MobileNet, DenseNet, and Capsule Network models on COVID-19 CT data.Three test lists contain approximately 11,450 CT scan slices.
- Benchmark evaluation: The Capsule Network achieves high sensitivity with lower specificity, while SegCaps-based Capsule Network achieves the best reported performance in Figure 9.The paper states that SegCaps provides the highest sensitivity and lowest specificity among the shown models.
- Federated-learning evaluation: Federated-learning accuracy changes as the number of hospitals or providers increases, while reported runtime varies by dataset and iteration count.The experiments distribute data across three hospitals to measure federated-learning performance.
- Benchmark evaluation: The Capsule Network exhibits the highest sensitivity, whereas ResNet has the best specificity in Table IV.These are the table’s stated best-performing models for the two measures.
- Federated-learning evaluation: Federated learning does not affect accuracy relative to the local model while providing privacy during data sharing.The local model uses the whole dataset, whereas federated learning combines locally trained models.
- Security analysis: The blockchain mechanism is described as decentralized, automated, traceable, and resistant to data tampering.Data providers retain control through ownership signatures and smart-contract policies.
CT Image U-NET U-NET ++ Ours
The paper compares its COVID-19 CT segmentation approach with established methods and related blockchain or federated-learning systems. Its method is reported to outperform baseline segmentation methods while achieving comparable federated-learning accuracy.
- CT Image U-NET U-NET ++ Ours: The visual comparison presents original CT images, overlay segmentation results, and segmentation masks for the proposed method and other studies.The first two columns show overlays, while the remaining columns show masks.
- Comparison with prior methods: Prior COVID-19 detection studies are described as not considering collaborative data sharing for training prediction models.The paper also contrasts these studies with methods using GANs or data augmentation, whose medical-image performance it calls unreliable.
- CT Image U-NET U-NET ++ Ours: The proposed infection segmentation outperforms baseline methods and produces results close to the ground truth.The paper states that U-Net++ performs near the proposed method.
- Comparison with prior methods: The federated-learning model achieves accuracy similar to training a local model on the whole dataset or combining weights across hospitals.The comparison includes VGG, ResNet, ImageNet, MobileNet, DenseNet, and Capsule Network models.
- Comparison with prior methods: The proposed framework differs from cited blockchain-based medical-image sharing by aggregating neural-network weights over a blockchain-enabled federated-learning system.The cited federated-learning frameworks addressed vehicle data rather than medical CT data.
IV. RELATED WORK
The related work positions AI and deep learning as tools for automated medical-image diagnosis, while federated learning enables collaborative training across hospitals with privacy concerns. The paper extends this direction with normalization, Capsule Network analysis, and blockchain-enabled federated learning for COVID-19 CT detection.
- AI and deep learning: Deep learning has reported high diagnostic accuracy for automated lung-disease detection in medical imaging.
- AI and deep learning: AI-based CT analysis tools have been developed to detect, quantify, and monitor coronavirus and distinguish infected from disease-free patients.
- Federated learning: Federated learning aggregates parameters from multiple organizations while allowing hospitals to share models instead of directly exposing their datasets.
- Proposed direction: The proposed framework addresses heterogeneous hospital data through normalization and combines Capsule Network segmentation and classification with blockchain-based federated learning.
- Proposed direction: The Capsule Network achieved the highest accuracy among the evaluated deep learning models, while hospitals shared private data to train a global model.