Diagnosing PSP Using the Hummingbird Sign and Parkinsonism Index in Deep Learning

Overview

Progressive Supranuclear Palsy (PSP) is a neurodegenerative disease where 40-60% of all PSP patients are initially misdiagnosed as having Parkinson's Disease due to identical physiological symptoms. This misdiagnosis causes patients to lose out on valuable treatment time, worsening their prognosis. One unique aspect of PSP however, is the Parkinsonism index. This index is a ratio of the area of the midbrain divided by the area of the pons, which can indicate whether an individual has Parkinson’s or PSP. Another unique aspect of PSP is the formation of the hummingbird sign, where neuron death atrophy due to excessive tau protein results in what is left of the pons, midbrain, and medulla oblongata to be shaped like a hummingbird. Unfortunately, both Parkinson’s disease and PSP can cause the hummingbird sign, but there are slight differences in their orientation and size. Research has found that each hummingbird sign of different diseases correlates to different species of hummingbirds. A patient with PSP would have the hummingbird sign in the shape of the hummingbird species the Rufous-Breasted Hermit (Glaucis Hirsutus). This project seeks to eliminate the misdiagnosis of PSP by creating a deep learning model to calculate the Parkinsonism index, and a CNN (Convolutional Neural Network) to identify the hummingbird sign in MRIs of PSP patients in order to test which biomarker is more accurate when used for diagnosis. Three separate models in python (using keras and tensorflow) were created: the first model (control) was a simple, 8-layer CNN, the second CNN model first trained on differentiating the hummingbird species before the MRIs, and the last deep learning model used a hybrid approach to take the areas of the midbrain/pons. With a sample size of 981 sagittal T1-weighted MRIs of PSP, Parkinson’s, and healthy patients, the models were run. Model 1 was quite inaccurate with a validation accuracy of 57.8%. As predicted, the deep learning model outperformed the second CNN with a validation accuracy of 94.33% while the validation accuracy for the second CNN was 93.1%. One limitation for this project was the smaller sample size, which can potentially be increased in the future. Eventually, these two models can be combined to create a new, more accurate diagnostic tool for PSP patients. Some future applications of this project could be using the combined model and training it with MRIs of other neurodegenerative diseases, one day making it able to diagnose other similarly misdiagnosed pathologies such as Multiple System Atrophy(MSA) and Alzheimer's Disease.

Competition history

• AJAS 2026 Category not listed

Related projects

Closest projects by meaning, across every fair and year in the corpus.

Browse more like this

• AJAS None projects
• Every AJAS 2026 project