Leeds Researchers Developed A Swarm Learning-Based AI Method To Predict Cancer From Patient Data Without Putting Personal Information At Risk
This Article Is Based On The Research Paper 'Swarm learning for decentralized artificial intelligence in cancer histopathology'. All Credit For This Research Goes To The Researchers Of This Paper 👏👏👏 Please Don't Forget To Join Our ML Subreddit
From standard histopathology slides, artificial intelligence (AI) can predict the presence of molecular changes. However, developing robust AI systems necessitates vast datasets, posing practical, ethical, and legal challenges. Swarm learning (SL) could help overcome these challenges by allowing partners to train AI models while eliminating data transmission and monopolistic data control.
A team of medical scientists has devised a novel technique to use artificial intelligence to forecast cancer from patient data without putting personal information in danger. Artificial intelligence (AI) can analyze enormous volumes of data, such as photographs or trial results, and spot patterns that humans can’t, making it extremely useful for disease detection, diagnosis, and treatment.
However, deploying the technology in medical settings is problematic because of the potential for unintentional data leaks. Various systems are owned and controlled by private corporations, giving them access to secure patient data — and the responsibility for securing it.
The researchers wanted to see if swarm learning, a type of AI, could help computers predict cancer in medical photos of patient tissue samples without disclosing the information from hospitals.
Swarm learning teaches AI algorithms to spot patterns in data from a local hospital or institution, such as genetic changes in human tissue photos. The swarm learning system then communicates this newly trained algorithm to a central computer, but not any local data or patient information. It is then merged with algorithms created in the same way by other hospitals to build an optimized algorithm. This is then returned to the local hospital, where it is reapplied to the original data, allowing for more sensitive detection of genetic mutations.
By repeating this process numerous times, the algorithm can be improved, and one that works with all data sets generated. This means the technology can be used without releasing any information to third parties or sending data between hospitals or international borders.
The researchers used data from three groups of patients from Northern Ireland, Germany, and the United States to train AI algorithms. The algorithms were tested on two vast sets of data images collected at Leeds, and they were found to predict the presence of several subtypes of cancer in the photos.
SL can be used to train AI pathology models
An SL-capable AI pipeline (Fig. 1a,b and Extended Data Fig. 1a–e) was established for molecular categorization of solid tumors based on histopathology pictures. Epi700 (n = 661 patients from Northern Ireland; Extended Data Fig. 2), DACHS (Darmkrebs: Chancen der Verhütung Durch Screening, n = 2,448 patients from southwestern Germany; Extended Data Fig. 3), and TCGA (The Cancer Genome Atlas, n = 632 patients Fig. 1c and Extended Data Fig. 4) were the three large datasets that were used for training.
BRAF mutation status can be predicted using SL models
In the QUASAR cohort (n = 1,774 patients from the United Kingdom – Data Fig. 5), patient-level performance was evaluated for predicting BRAF mutational status. When trained just on Epi700, DACHS, and TCGA, local models obtained areas under the receiver operating curve (AUROCs; mean sd) of 0.7358 0.0162, 0.7339 0.0107, and 0.7071 0.0243, respectively (Fig. 2a). The prediction AUROC was increased to 0.7567 0.0139 by combining the three training cohorts on a central server (merged model) (P = 0.0727 vs Epi700, P = 0.0198 vs DACHS, P = 0.0043 vs TCGA; Fig. 2a).
Microsatellite instability can be predicted using SL models
Next, testing the prediction pipeline by predicting microsatellite instability (MSI)/mismatch repair deficiency (dMMR) status in the QUASAR clinical trial cohort (Fig. 2b) and the YCR BCIP population-based cohort (Yorkshire Cancer Research Bowel Cancer Improvement Programme; Fig 2c and Extended Data Fig. 6). In QUASAR, b-chkpt1 and b-chkpt2 had prediction AUROCs of 0.8001 0.0073 and 0.8151 0.0071, respectively, outperforming single-cohort models trained on Epi700, which had an AUROC of 0.7884 0.0043 (P = 0.0154 and P = 8.79 105, respectively) .
The data efficiency of SL models is impressive
In medical AI, learning from small datasets is difficult since prediction performance improves as the size of the training dataset grows. As a result, it was checked whether SL could compensate for the loss of performance when just a small selection of patients from each institution is trained. For single-dataset (local) models, limiting the number of patients in each training set to 400, 300, 200, and 100 significantly reduced prediction performance. Training on only a subset of patients in Epi700, DACHS, or TCGA, for example, significantly lowered prediction performance and increased model instability in QUASAR, as indicated by a greater interquartile range of predictions in experimental repetitions (Fig. 3a).
Models that learn plausible patterns are known as SL models
Medical AI models should perform well and be easy to understand. The model evaluated predictions on a millimeter-scale (Fig. 2d). These maps revealed a clear and consistent superiority of one of the classes. Furthermore, the model assessed predictions on a micrometer scale by extracting the image patches with the highest scores for models trained on 300 patients and all patients from the local training cohorts (Fig. 4a–c), the merged cohort (Fig. 4d), and the swarm models b-chkpt1, b-chkpt2, and w-chkpt1 (Fig. 4e,f).
The SL principle
The idea of SL is to train a machine learning model in many computer systems that are physically separated. In this case, SL is used in a network of three physically distinct machines (peers). After each sync period, multiple synchronization (sync) events send model weights from each partner to the other peers. At each sync event, model weights are averaged before training resumes using the averaged parameters at each peer. Unlike FL, there is no central instance that merges the parameters regularly. Instead, smart contracts on the Ethereum blockchain allow the network to choose any peers to combine parameters at each sync stop. The blockchain keeps track of the model’s global state in this configuration. Two varieties of SL were created, tested, and evaluated: basic and weighted.