Cancer is the most deadly and dreaded disease ever encountered by mankind and tumor size plays a crucial role in determining the severity and treatment for the same. Therefore, it becomes imperative to estimate the dimensions of the associated tumor with paramount accuracy and precision so as to enable radiologists and doctors, in general, to effectively prescribe a treatment post-diagnosis. Current estimation approaches of tumor size involve the manual click and drag measurements by radiologists which are functional but prone to a lot of manual errors and redundancies. To improve the overall accuracy and efficiency of the process, the authors propose a Deep learning solution that uses DICOM scan images to determine the dimensions of the tumor. Furthermore, this solution provides a 3D representation of the tumor for clear perception and comprehension and also provides treatment suggestions that aid doctors throughout the treatment. The pipeline consists of two models namely, CNN model for detection performs with an accuracy of 97.6% and a ResUNet model to segment tumor out of the brain image with accuracy of 91.54%.

Background: Deep Learning is a subset of Artificial In- telligence (AI) and Machine Learning (ML) which can understand and learn hidden features in images and make better predictions. Healthcare field is increasingly seeing automated AI involvement in making disease predictions by providing sufficient labeled data. Alzheimer’s disease is a neurological condition in which the death of brain cells causes memory loss and cognitive decline. Objective: Currently, most Deep Learning models for Alzheimer’s Dis- ease prediction uses cloud server architecture which increases the infer- ence time manifold and opens up the risk of data breach as personal data of subjects are sent to the cloud for computation. This work aims to de- tect Alzheimer’s Disease and deploy it on a scalable edge-device such as Raspberry Pi. Methods: A weighted ensemble network is proposed to serve the pur- pose. This method includes individual models such as AlexNet, DenseNet- 121, and InceptionV3. The ensemble network is build atop these networks and selects the best performing model for a given input image. Results: The accuracy and sensitivity obtained using the proposed weighted ensemble network is 98.44% and 98.44% respectively. The inference time of classifying an input recorded on Raspberry Pi 4 Model B is 1.97 sec- onds. The weighted ensemble network produced 5% and 19% improve- ment in accuracy and 6.5% increase in sensitivity when comparing with related research work in the domain. Conclusion: From this work, it can be seen that there is a large poten- tial to make Alzheimer’s Disease Detection scalable on edge-devices and making it real-time.