Research On Craniopharyngioma Invasiveness Diagnosis Based On Modified CNN Algorithm

Craniopharyngioma is a kind of benign brain tumor,but its invasion of the surrounding brain tissue can cause severe manifestations,such as hypothalamic-pituitary dysfunction,increased intracranial pressure,and visual impairment.Due to the similarity between invasive and non-invasive craniopharyngioma lesions,it is considered a challenge for radiologists to diagnose the invasiveness of craniopharyngioma through preoperative MRI images.Therefore,developing a computer aided diagnosis system that can predict the invasiveness of craniopharyngioma through preoperative MRI images is of clinical value for making more appropriate and individualized treatment plans.By using deep learning algorithms,computer aided diagnosis systems are proposed for diagnosing the invasiveness of craniopharyngioma in this thesis.In summary,the main contents of this thesis are as follows:(1)Construct an MRI dataset of craniopharyngioma and data preprocessing.Those preoperative MRI images and pathological results of 576 patients with craniopharyngioma(196 patients with invasive and 380 patients with non-invasive craniopharyngioma)were enrolled in this thesis.The ITK-SNAP was used for MRI images annotation.So far,the MRI image database of craniopharyngioma has been constructed.In order to remove the background and augment the database during data preprocessing,the processes including brain cropping,contrast adjustment,random rotation,and zooming were designed.(2)Build the context aggregation(CA)deep learning network.The CA network is a lightweight model based on 2D CNN.The residual network is utilized as the backbone for extracting features.The context modeling block and the feature aggregation block based on the attention mechanism are introduced in the CA network for capturing the context information between different MRI slices and aggregating features.The diagnosis performance is improved through those changes.(3)Build the multi-task two-path deep learning system for brain tumors(MT-Brain).The MT-system is composed of a sub-path based on 2D CNN and another subpath based on 3D CNN,which are used to extract 2D features from each slice of MRI images and additional 3D spatial context features between MRI slices.In addition,the position encoding block and maskguided attention block are also introduced.Finally,the MT-Brain can locate the craniopharyngioma lesion and diagnose the invasiveness of craniopharyngioma synchronously.(4)Explore the interpretability of deep learning models.Focusing on the uninterpretability of deep learning models,the t-SNE algorithm and feature heat maps are used in this thesis to prove the strong discriminative ability of the proposed models.Besides,the Grad-CAM algorithm is used to visualize the brain area that draws the most attention and prove the rationality of the models.