HOẠT ĐỘNG TRONG TUẦN

Tóm tắt: Our research aimed to predict the survival of non-small cell lung cancer (NSCLC) patients undergoing immunotherapy within 12 months using 18F-FDG PET images. We developed machine learning and deep learning models and evaluated their performance on various data types, including structured data such as radiomics features (surrogate total metabolic tumor volume (sTMTV), surrogate tumor dissemination (sDmax)), and clinical data (type of cancer, smoking status, patient status, age, gender, etc), as well as unstructured data such as 2D images. Several machine learning models, including k-nearest neighbors (KNN), Support Vector Machine (SVM), and Random Forest (RF), were trained on the structured data using the 5-fold cross-validation technique. Besides, the DenseNet121 model was also fine-tuned separately on three different types of images, including whole 2D maximum intensity projection (MIP) images, 2D mask regions, and 2D segmented images, using the same 5-fold cross-validation technique, where the patients in each fold are the same as for the machine learning models. Our study produced promising results, indicating that sTMTV is a valuable feature for predicting NSCLC patients' survival, consistent with previous literature. We also found that incorporating lesion segmentation in the deep learning model reduced the number of features that needed to be learned, improving the overall model performance. Moreover, our study demonstrated the vital role of clinical data in accurately predicting patient survival. By combining clinical data with radiomics and CNN features, our machine learning models can achieve a significant boost in performance.