01Research problem
Early in the pandemic, test kits were scarce and not always reliable, and CT imaging offered a complementary route to diagnosis. Deep-learning detectors need far more data than any single hospital had, yet hospitals could not pool patient scans because of privacy rules. Scans from different hospitals also came from different CT scanners, so the data were heterogeneous as well as scarce.
02What is new
- Normalisation for multi-hospital, multi-scanner data, so models can learn from heterogeneous CT sources.
- Capsule-network detection: SegCaps segmentation followed by Capsule Network classification of COVID-19 patterns in lung CT.
- Blockchain-based federated learning: hospitals share locally trained model weights, not scans; a blockchain authenticates contributions and Laplace noise adds differential privacy.
- A new public dataset, CC-19, released to the research community.
03Method
Data. CC-19 contains 34,006 CT slices (28,395 COVID-positive) from 89 subjects, 68 positive and 21 negative, collected at three hospitals in Chengdu on six scanners, about 231 CT volumes in total.
- Normalise slices using Hounsfield-unit ranges to reduce scanner differences.
- Segment lungs and infection with SegCaps, then classify with a Capsule Network.
- Train local models at each hospital, adding Laplace noise for differential privacy.
- Aggregate local models into a global model through blockchain consensus, which also checks local model quality.
Baselines. Thirteen ImageNet-pretrained networks (VGG, AlexNet, Xception, ResNet, Inception, MobileNet, DenseNet) were fine-tuned with Adam at a learning rate of 10−5 with early stopping.
04Results
| Network | Pre-trained | Precision | Sensitivity / recall |
|---|---|---|---|
| VGG16 | ImageNet | 0.827 | 0.829 |
| ResNet50 | ImageNet | 0.833 | 0.771 |
| DenseNet121 | ImageNet | 0.832 | 0.903 |
| MobileNet | ImageNet | 0.830 | 0.912 |
| SegCaps + Capsule Network (ours) | – | 0.830 | 0.987 |
Federated training. Distributing the data over three hospital nodes reached accuracy similar to training one model on all the data, while keeping each hospital's scans local. Running cost grows with the number of hospitals and transactions.
Sensitivity was prioritised because missing a COVID-19 case is costly; precision is comparable to the baselines rather than higher. The evaluation is slice-level, on data from hospitals in one city.
05Reproducibility
| Item | Details | Status |
|---|---|---|
| Paper | Open access in PubMed Central (PMC8791443) | Available |
| CC-19 dataset | Public on GitHub, hosted by co-author Abdullah Aman Khan; no explicit licence file | Available, unlicensed |
| Training and federation code | Not publicly released | Not public |
| Hyperparameters | Optimiser, learning rate, decay and early stopping reported for the baselines | Partly reported |
06My contribution
First author of nine: Rajesh Kumar, Abdullah Aman Khan, Jay Kumar, Zakria, Noorbakhsh Amiri Golilarz, Simin Zhang, Yang Ting, Chengyu Zheng and Wenyong Wang (University of Electronic Science and Technology of China).
07Limitations
- Small cohort (89 subjects) from one region; no external multi-site validation.
- Slice-level evaluation rather than patient-level diagnosis.
- Training and federation code are not public, which limits independent reproduction.
- Developed in the early-pandemic context of 2020–2021.
08Cite
@article{kumar2021blockchain,
title = {Blockchain-Federated-Learning and Deep Learning Models for COVID-19 Detection Using CT Imaging},
author = {Kumar, Rajesh and Khan, Abdullah Aman and Kumar, Jay and Zakria and Golilarz, Noorbakhsh Amiri
and Zhang, Simin and Ting, Yang and Zheng, Chengyu and Wang, Wenyong},
journal = {IEEE Sensors Journal},
volume = {21},
number = {14},
pages = {16301--16314},
year = {2021},
doi = {10.1109/JSEN.2021.3076767}
}A related real-time CT analysis demo, built on this line of blockchain-secured medical-imaging work, is shown below.