Gliomas are the most common primary malignant brain tumors in adults. Accurate grading is crucial as therapeutic strategies are often disparate for different grades and may influence patient prognosis. This study aims to provide an automated glioma grading platform on the basis of machine learning models. In this paper, we investigate contributions of multi-parameters from multimodal data including imaging parameters or features from the Whole Slide images (WSI) and the proliferation marker Ki-67 for automated brain tumor grading. For each WSI, we extract both visual parameters such as morphology parameters and sub-visual parameters including first-order and second-order features. On the basis of machine learning models, our platform classifies gliomas into grades II, III, and IV. Furthermore, we quantitatively interpret and reveal the important parameters contributing to grading with the Local Interpretable Model-Agnostic Explanations (LIME) algorithm. The quantitative analysis and explanation may assist clinicians to better understand the disease and accordingly to choose optimal treatments for improving clinical outcomes. The performance of our grading model was evaluated with cross-validation, which randomly divided the patients into non-overlapping training and testing sets and repeatedly validated the model on the different testing sets. The primary results indicated that this modular platform approach achieved the highest grading accuracy of 0.90 ± 0.04 with support vector machine (SVM) algorithm, with grading accuracies of 0.91 ± 0.08, 0.90 ± 0.08, and 0.90 ± 0.07 for grade II, III, and IV gliomas, respectively.
18F-FDG PET / CT is used clinically for the detection of extramedullary lesions in patients with relapsed acute leukemia (AL). However, the visual analysis of 18F-FDG diffuse bone marrow uptake in detecting bone marrow involvement (BMI) in routine clinical practice is still challenging. This study aims to improve the diagnostic performance of 18F-FDG PET/CT in detecting BMI for patients with suspected relapsed AL. Methods: Forty-one patients (35 in training group and 6 in independent validation group) with suspected relapsed AL were retrospectively included in this study. All patients underwent both bone marrow biopsy (BMB) and 18F-FDG PET/CT within one week. The BMB results were used as the gold standard or real "truth" for BMI. The bone marrow 18F-FDG uptake was visually diagnosed as positive and negative by three nuclear medicine physicians. The skeletal volumes of interest were manually drawn on PET/CT images. A total of 781 PET and 1045 CT radiomic features were automatically extracted to provide a more comprehensive understanding of the embedded pattern. To select the most important and predictive features, an unsupervised consensus clustering method was first performed to analyze the feature correlations and then used to guide a random forest supervised machine learning model for feature importance analysis. Cross-validation and independent validation were conducted to justify the performance of our model. Results: The training group involved 16 BMB positive and 19 BMB negative patients. Based on the visual analysis of 18F-FDG PET, 3 patients had focal uptake, 8 patients had normal uptake, and 24 patients had diffuse uptake. The sensitivity, specificity, and accuracy of visual analysis for BMI diagnosis were 62.5%, 73.7%, and 68.6%, respectively. With the cross-validation on the training group, the machine learning model correctly predicted 31 patients in BMI. The sensitivity, specificity, and accuracy of the machine learning model in BMI detection were 87.5%, 89.5%, and 88.6%, respectively, significantly higher than the ones in visual analysis (P < 0.05). The evaluation on the independent validation group showed that the machine learning model could achieve 83.3% accuracy. Conclusions: 18F-FDG PET/CT radiomic analysis with machine learning model provided a quantitative, objective and efficient mechanism for identifying BMI in the patients with suspected relapsed AL. It is suggested in particular for the diagnosis of BMI in the patients with 18F-FDG diffuse uptake patterns.
Medical imaging is playing an increasingly crucial role in disease diagnosis. Numerous deep learning‐based methods have been developed for image‐guided automatic disease diagnosis. Most of the methods have harnessed conventional convolutional neural networks, which are directly applied in the regular image domain. However, some irregular spatial patterns revealed in medical images are also critical to disease diagnosis, since they can describe latent relations in different image regions of a subject (e.g., different focal lesions in an image) or between different groups (e.g., Alzheimer's disease and healthy control). Therefore, how to exploit and analyze irregular spatial patterns and their relations has become a research challenge in the field of image‐guided disease diagnosis. To address this challenge, graph neural networks (GNNs) are proposed to perform the convolution operation on graphs. Graphs can naturally represent irregular spatial structures. Because of their ability to aggregate node features, edge features, and graph structure information to capture and learn hidden spatial patterns in irregular structures, GNN‐based algorithms have achieved promising results in the detection of various diseases. In this paper, we introduce commonly used GNN‐based algorithms and systematically review their applications to disease diagnosis. We summarize the workflow of GNN‐based applications in disease diagnosis, ranging from localizing the regions of interest and edge construction to modeling. Furthermore, we discuss the limitations and outline potential research directions for GNNs in disease diagnosis.