As of 2017, NIH's deep learning system matching semantic labels (i.e.: diagnoses) to images had a top-one & top-five accuracy of 61-66% & 93-95%.
How is a convoluted neural network classifying radiology images assessed ?
1. Accuracy = Corrected predicted samples/All predictions
2. Top-five accuracy (assessing if the correct label pertains to the five highest predicted classes)
3. Confusion matrix - predicted v. true labels
Employing CPUs vs. GPUs for deep learning in radiology
CPUs
- Large memory size
- Used for sequential processing
- Restricted memory bandwidth
GPUs
- Limited memory size
- Used for parallel processing
- High bandwidth
GPUs are preferred over CPUs.
4. Monitoring (i.e.: longitudinally determine regression or stagnation of lesions on imaging, reocurrence of arrhythmias such as non-sustained VT)
5. Prognosis (e.g.: overall survival in cancer patients; burden and arrhythmia-free interval)
Deep learning and its healthcare applications
The AI tasks used in healthcare are:
1. Preprocessing (e.g.: artifact reduction on CT scans/EKGs)
2. Detection (boxing lesions on imaging, pattern recognition of arrhythmias)
3. Classification (binary - benign/malignant; multiclass)
New Research: Thyroid-related adverse events induced by immune checkpoint inhibitors: Immune checkpoint inhibitors, namely anti-CTLA-4, anti-PD-1 and anti-PD-L1 monoclonal antibodies, have emerged in the last decade as a novel form… https://t.co/Mskq56ezlx #Endocrinology#Endo