December 9, 2023
By Mark L. Fuerst
A novel artificial intelligence (AI) model can quickly distinguish between prefibrotic primary myelofibrosis (PMF) and essential thrombocythemia (ET) with high accuracy in distinct clinical cohorts. The model, previously trained on more than 32,000 pan-cancer biopsy images that has learned general pathological features, is the largest test to use AI to differentiate between prefibrotic PMF and ET. The AI-based system used patient images from the U.S. and Italy and was able to return results in just over 6 seconds for a new patient, on average, with an overall accuracy of 92.3 percent.
Myeloproliferative neoplasms are a family of hematological myeloid neoplasms. In 2016, the World Health Organization differentiated ET from newly defined prefibrotic PMF. Compared to ET, prefibrotic PMF patients have a higher risk for constitutional symptoms, major hemorrhage, and progression to overt MF and leukemia, lead author Andrew Srisuwananukorn, MD, Assistant Professor at The Ohio State University Comprehensive Cancer Center, told a press conference at the 2023 American Society of Hematology annual meeting (Abstract 901). Also, the median overall survival significantly differs between the two disorders—11.9 years for prefibrotic PMF and 22.2 years for ET.
“The diagnosis of ET and prefibrotic PMF rely on similar clinical analysis, mutational profiling, morphological analysis of megakaryocytic proliferation and atypia, and the subjective assignment of a fibrosis grade by a hematopathologist,” Srisuwananukorn said. There is high interobserver variability in bone marrow assessment with consensus varying between 50 percent and 100 percent. “There is a pressing need for improved diagnostics to differentiate among prefibrotic PMF and ET,” he said.
In a multicenter study, researchers set out to develop and validate a biologically motivated AI algorithm to rapidly, accurately, and inexpensively diagnose prefibrotic PMF and ET directly from diagnostic bone marrow biopsy digital whole-slide images.
“To our knowledge, this study represents the largest image-based AI study within MPNs with external validation,” Srisuwananukorn noted. “Our proposed model may assist clinicians in appropriately identifying patient cohorts who would benefit from disease-specific therapies or enrollment in clinical trials. We imagine that a potential high-speed, low-cost algorithm may reliably distinguish prefibrotic PMF from ET patients with high specificity, which can be democratized to the MPN clinical community in routine practice and drive clinical trial accrual for biologically rational novel therapeutics.”
Patients with a clinical/histopathological diagnosis of prefibrotic PMF or ET were identified at the University of Florence in Italy between June 2007 and May 2023 and at Moffitt Cancer Center in Tampa, FL, between January 2013 and January 2022. The training cohort comprised of 200 (100 prefibrotic PMF/100 ET) patients from Florence and the external test cohort entailed 26 (6 prefibrotic PMF/20 ET) patients from Moffitt.
In total, the resultant model was trained on 32,226 patient-derived whole-slide images. A prediction upon each patient’s whole-slide image was calculated by attention-based multiple instance learning, which is a method that automatically assigns a numeric weight to an image portion representing its relative importance to the classification task.
Within the training cohort, 5-fold cross-validation resulted in a mean AUC of 0.90 and a standard deviation of 0.04. A final locked model re-trained on the entire training cohort resulted in an AUC of 0.90 upon evaluation of the test cohort. The final diagnostic classification accuracy on the test cohort was 92.3 percent with a sensitivity and specificity for prefibrotic PMF diagnosis of 66.6 percent and 100 percent, respectively.
“With the combined accuracy, sensitivity, and specificity we saw, it would allow the physician to be confident in one diagnosis versus another and help rule in or rule out the rarer prefibrotic PMF diagnosis, particularly for clinical trials,” Srisuwananukorn said. “My hope is that it would maintain this accuracy when tested in larger cohorts.”
Upon review of the slides with the highest prediction value per class, attention heatmaps highlighted the model’s reliance on areas of cellular marrow without reliance on image artifacts or background. Using affordable consumer-grade hardware, evaluation upon a previously unseen whole-slide image was completed in approximately 6.1 seconds.
Srisuwananukorn said the algorithm could potentially be used as a companion tool for clinical diagnoses and help doctors match patients with the clinical trials that are most likely to help them, which could ultimately result in better treatments. The algorithm is intended to complement, not replace, human experts.
“What we’re trying to develop is a clinical decision support tool with an emphasis on support,” said Srisuwananukorn. “Physicians with no computer science background are increasingly recognizing the value of AI algorithms and closer to being able to use them for their clinical practice. However, more investigations would be needed for this algorithm to be used in clinical practice, including testing in cohorts with different racial backgrounds.”
The researchers plan to continue to refine the system and test it with larger data sets. In addition, Srisuwananukorn said AI could potentially be used to advance basic research on MPNs to link biological processes with particular morphological features visible on biopsy slides. Eventually, this could lead to ways to predict a person’s prognosis or response to treatment based on biopsy images. Ongoing evaluations are seeking higher resolution with the AI-based system, which is now available through an open-source package.
Mark L. Fuerst is a contributing writer