Artificial intelligence based malignant lymphoma type prediction using enhanced super resolution image and hybrid feature extraction algorithm.
Journal:
Computers in biology and medicine
Published Date:
May 31, 2025
Abstract
In the medical field, the most common and frequent type of blood cancer is lymphoma. Accurately predicting and early response to lymphoma treatment will be useful for initiating treatment plans to achieve a greater rate of cure or reduced risk of treatment-related morbidity and death. Numerous research studies have revealed several pre-treatment predictive variables, including age, lactate dehydrogenase (LDH), Ann Arbor stage, and ECOG performance. The 18F-FDG PET imaging method makes it possible to assess human tissues' glycolytic activity at the cellular level. It has been extensively utilized for lymphoma staging and diagnosis. Its usefulness for evaluating the response to therapy in lymphoma patients has been demonstrated in a number of promising investigations. However, target training data are typically insufficient for deep neural network training in outcome prediction tasks, making it challenging to provide accurate prediction. Thus, the ensemble learning based target training model has been used with the hybrid feature extraction models, to train the model based on texture, colour and shape. The various types of malignant lymphoma images are gathered as a dataset and pre-processed using spatial image filtering procedures to remove specific unwanted spatial frequencies from an image. Contrast stretching is a technique used to enhance contrast in images and the NNU-Net has been employed to attain the super resolution images for learning in-depth information from the images. After pre-processing the images, the GLRLM, CLCM and HU-moment features extraction techniques are used for extracting the texture, shape and colour features from the images. In order to forecast the type of malignant lymphoma, the retrieved images are finally classified utilizing the Bi-LSTM, DBN, and RBFN by stacking and providing the gathered learning features from each algorithm to the Meta classifier. The proposed model's performance is compared with the most cutting-edge transfer learning techniques and validated using measures including classification Accuracy, Positive predictive value, Hit rate, Selectivity and NPV, whose values are 94.8 %, 91.0 %, 90.6 %, 93.30 % and 95.6 %. Thus, a malignant lymphoma type prediction based on an enhanced super resolution image and ensemble learning classifier outperforms the existing model.