There are many active research projects accessing and applying shared ADNI data. Use the search above to find specific research focuses on the active ADNI investigations. This information is requested annually as a requirement for data access.
| Principal Investigator | |
| Principal Investigator's Name: | Anish Salvi |
| Institution: | ImageRx |
| Department: | Student-Led Medical Imaging Software Company |
| Country: | |
| Proposed Analysis: | Unlocking information confined within terabytes of images, the study incorporates recent deep-learning advances from unique network designs to regional attention-based image transformers. Our three specific aims include utilizing AI to: 1) classify T1 MRIs as AD or normal control (NC) via a 3D residual neural network (ResNet), 2) train a 3D convolutional neural network (U-Net) to segment the MTL from TI MRIs and conduct shape analysis relative to a normal brain atlas to discern quantifiable departures from normalcy, and 3) classify AD with regional attention to the MTL evidenced by the neural embeddings (compressed feature representations) of the pretrained MTL segmentation model. The T1 image cohort will be interpolated to a standard volume and normalized before being organized into 90% training and 10% testing with stratified sampling. Aim 1: A 3D ResNet will be trained to predict AD status (AD vs NC). During previous experiments, it was noted that learning rate, batch size, and epochs significantly influenced model performance. Hyperparameter optimization will be conducted per 5-fold cross validation with stratified sampling. The success criteria is defined as accuracy, sensitivity, and specificity of >0.95 in the test set. Aim 2: A 3D U-Net will be trained to automatically segment the MTL volume in high-resolution by using localization as a prerequisite for segmentation. For the localization step, the 3D U-Net will be trained on the clinically acquired T1 MRIs and MTL ground truths. The predicted segmentations will then be resampled to their clinically acquired MRI counterparts while preserving the internal spacing, origin, and units. The clinically consistent predictions will be binarized via an Otsu threshold to separate MTL foreground from cerebral tissue background. Fit to the predicted segmentations, bounding boxes will be applied to the clinically acquired T1 MRIs. The Intersection-over-Unions comparing ground truth to predicted bounding boxes will be calculated based on image-specific coordinates with a success criteria of >0.5 in the test set following 5-fold cross validation. For the high-resolution segmentation step, the 3D U-Net will be trained on the cropped T1 MRIs and ground truths after standardizing image volumes and normalization. The same post-processing steps will be applied to the localized predicted segmentations which will be reinserted into the clinically consistent T1 MRIs frames of reference. The Dice Score Coefficients will be tabulated with a success criteria of >0.9 in the test set following 5-fold cross validation. Shape analysis will be conducted to view geometrical differences from normalcy. Aim 3: The localized pretrained 3D U-Net segmentation model will be leveraged as an image transformer which converts 3D T1 MRIs into neural embeddings. During segmentation model inference, images are passed through the 3D U-Net’s contracting path, culminating as feature maps within the bottleneck layer. We refer to these feature maps as neural embeddings. The neural embeddings will be used to train the same 3D ResNet for classification of AD status. Hyperparameter optimization will be conducted per 5-fold cross validation with stratified sampling. As before, the success criteria is defined as accuracy, sensitivity, and specificity of >0.95 in the test set. |
| Additional Investigators | |
| Investigator's Name: | Prahlad G. Menon |
| Proposed Analysis: | AD is a neurodegenerative disorder rising in prevalence as the elderly population grows. From 2021 to 2050, the number of people >65 affected by AD is estimated to rise from 6.2 to 13 million. During the same period, the cost of AD management is projected to rise from $355 billion to $1.1 trillion. The affliction progresses from forgetfulness to despondence, culminating in the failure of bodily function due to severe brain atrophy. Early diagnosis and intervention are critical for delaying Alzheimer’s symptoms and ensuring significant savings for the healthcare system. Medical imaging offers insight into a patient’s AD status. CT and MRI generate detailed anatomical pictures of the brain, allowing staff to view brain structure degradation, an indicator of AD. Functional imaging like PET scans identify amyloid-plaques and tau-proteins, biomarkers of AD. However, the burden of image interpretation falls upon radiologists. With a more than 10-fold surge in image processing, radiologists can suffer from fatigue, increasing the likelihood of errors. Burnout is especially prevalent amongst diagnostic radiologists and shortages of image interpreters is already leading to radiology departments falling behind on reports. Overwhelming radiologists, the incoming deluge of medical images will cause an increase in delayed and erroneous identification of AD status. Suffering from poorer health outcomes, the aging population will bear the brunt of misdiagnoses. The literature indicates that amyloid and tau accumulations, AD biomarkers present years before symptoms, are associated with atrophy of the medial temporal lobe (MTL), an early hallmark of AD. Structural imaging, such as T1 MRI, can be part of standard-of-care imaging workups to discern AD from other afflictions. Leveraging deep learning, our three specific aims are as follows: 1) classify T1 MRIs as AD or normal using a 3D residual neural network underpinned on skip connections, 2) train a neural network to segment the MTL from TI MRIs and subsequently conduct shape analysis relative to a normal brain atlas to discern quantifiable departures from normalcy, 3) classify AD with regional attention to the MTL evidenced by the neural embeddings (compressed feature representations) of the pretrained MTL segmentation model. Models will be optimized via 5-fold cross-validation and compression/quantization will be explored to reduce inference-related image processing times. Powered by AI analysis of medical images, ImageRx will improve early diagnosis and augment the radiology workflow through rapid and automated prediction of AD via T1 weighted MRI interpretation and MTL quantification. |
