Page 89 - AIH-2-4

Artificial Intelligence in Health Federated learning health stack against pandemics

reported for RT-PCR tests. A more detailed analysis architectures, including ResNet50, SeResNext50, 31,32
30
revealed that EfficientNet exhibited the highest specificity and DenseNet161, were evaluated, with DenseNet161
33
(95.4%) for COVID-19, while ResNet34 showed the highest demonstrating the optimal performance. The CNN classifies
sensitivity (94%). Interestingly, EfficientNet showed each ROI as cancerous or non-cancerous, providing not
the highest performance in classifying normal cases, only diagnostic information but also, combined with the
SeResNext50 excelled in classifying pneumonia cases, and clustering results, an approximate segmentation mask. This
ResNet34 was most effective for COVID-19 cases. These approach bypasses the need for computationally intensive
results are comparable to, and in some cases outperform, semantic segmentation techniques, such as GrabCut or
46
those reported in earlier studies using CNNs for COVID-19 context-aware masks, offering a more efficient method
47
detection from chest X-rays and computed tomography for identifying and localizing cancerous nodules within
(CT) scans. 37-39 The present deep-learning approach offers the lung.
advantages in terms of speed, resource efficiency, and
potentially greater accuracy compared to RT-PCR. The 2.8. Public data authentication and blockchain
approach is particularly appealing for resource-limited The previous section discussed the need to ensure a smooth
settings where rapid screening is critical. Moreover, the communication protocol between the central server, local
decoupled workflow, separating image acquisition and servers, and data-owning entities, which necessitates
diagnostic evaluation, offers greater operational flexibility entering into a legal contract between the parties. It is
and scalability, aligning with the principles outlined in the important to note that, ideally, the central server should be
previous study for building a robust pandemic response. owned either directly by the United Nations or any of its
29
A previous study explored a novel approach for subsidiaries, such as the WHO. Due to the lack of specific
40
lung cancer detection, combining an unsupervised sovereign oversight over the entire protocol, creating digital
clustering algorithm for ROI proposals with CNNs. 41,42 trust is imperative. Fortunately, a robust technical solution
The modularity optimization-based graph clustering is available to build trust in the digital space between
method 41,42 applied to preprocessed CT scans from the unknown agencies through a decentralized protocol called
2016 lung nodule analysis dataset 43,44 reduces CNN blockchain. Blockchain-based smart contracts offer an
complexity by identifying potential nodule locations. attractive method that creates a transparent technological
This preprocessing step includes lung segmentation using framework governing the relations between the parties
marker-controlled “watershed segmentation” on Sobel- participating in the FL architecture. However, such smart
filtered images, focusing the analysis on relevant areas and contracts are not free from threats, and in the following, a
reducing the computational burden. The segmented lung few novel methods to ensure data safety in smart contracts
regions are then converted into a network, with pixels are discussed.
representing nodes and edges connecting neighboring
pixels. This network is then clustered using a modularity Detecting vulnerabilities in smart contracts is crucial
function optimized for spatially embedded networks, 41,42 due to their immutable nature and the potential for
which has been successfully used previously in analyzing significant financial losses, as evidenced by past incidents,
48
granular assemblies. This method effectively identifies including the decentralized autonomous organization
43
49
50
nodules based on grayscale-intensity similarity, generating hack and the Parity wallet freeze. Gogineni et al.
ROI proposals for subsequent CNN analysis. constructed a multiclass classifier to detect vulnerabilities
in smart contracts by fine-tuning an AWD-LSTM model.
51
The proposed method utilizes these ROI proposals The model was fine-tuned using a dataset of 40,877
to streamline the CNN classification process. Rather unique opcode combinations from smart contracts.
than relying on computationally expensive selective The smart contracts were classified into four categories:
search or fully labeled datasets required for methods suicidal, prodigal, greedy, and normal. To address the
43
like MaskRCNN, this approach employs the clustering class imbalance, only distinct opcode combinations were
45
algorithm to generate a manageable number of fixed-size retained for normal smart contracts, as they comprised
ROIs. These ROIs are then fed into a CNN, trained using the majority of the dataset and often contained repeated
transfer learning initialized with ImageNet weights, and sequences.
35
optimized using learning rate scheduler techniques.
36
This training strategy, employing discriminative learning The AWD-LSTM model architecture combined a
rates for different layers of the network, facilitates efficient pretrained encoder with a custom classification head,
learning of both general and dataset-specific features, inspired by the ULMFiT method used in natural
52
which are of great value for the FL learning protocols language processing. This method achieved an overall
discussed in this article. Several pre-trained CNN accuracy of 91.3% and a weighted average F1 score of
Volume 2 Issue 4 (2025) 83 doi: 10.36922/AIH025080013