Integrating Adaptive Gaussian Precision and Multi-Scale Feature Extraction for Enhanced Diagnosis of Ph-Negative Myeloproliferative Neoplasms in BMT Images

Abstract

Philadelphia Chromosome (Ph)-negative myeloproliferative neoplasms (MPNs) are a critical subgroup of hematologic malignancies characterized by uncontrolled proliferation of myeloid cell lines in the bone marrow, leading to subtypes such as essential thrombocythemia (ET), polycythemia vera (PV), and primary myelofibrosis (MF). The study utilizes the Ph-negative MPN Image Dataset, an open-access resource of 300 bone marrow trephine (BMT) images divided into three classes. To address diagnostic challenges, a novel approach integrating six pre-trained convolutional neural network (CNN) models - VGG-19, DenseNet-201, Xception, InceptionV3, EfficientNetB7, and NasNetLarge - alongside an Adaptive Gaussian Precision Capture (AGPC) block is proposed. The backbone models extract fundamental cellular and stromal features, while the AGPC block enhances these by capturing multi-scale structural details, even in low-contrast areas. Key advantages of the AGPC block include its ability to utilize multi-scale Gaussian filters for simultaneous extraction of fine and broad structural features, suppression of background noise through its gamma parameter, and improved focus on diagnostically critical regions using adaptive attention maps. Preliminary experiments show that the AGPC block significantly improves EfficientNetB7's specificity (73.33% to 85.83%) and DenseNet-201's accuracy (86.67% to 90.00%), demonstrating its effectiveness in enhancing feature extraction. Notably, NasNetLarge remains the most efficient model, maintaining its high performance with or without the AGPC block, showcasing the differential impact of the integration. In conclusion, this study uses an innovative, AI-based approach to diagnose Ph-negative MPNs. It combines CNN models and AGPC blocks to improve accuracy. Further optimization could improve the approach, including hyperparameter and dataset changes, and integrating AGPC blocks with other architectures or data types would enhance its use. © 2025 IEEE.