Submission
| Title: | Applications of Machine Learning in Tissue Image Analysis |
| Presenter: | Annabelle White |
| Institution: | Whiteland Community |
| Authors: | Annabelle White, Whiteland Community High School, Whiteland, IN; Takashi Hato, PhD, Hato Lab, IU School of Medicine (Nephrology Division), Indiana University, Indianapolis, IN |
Abstract
| Background/Significance/Rationale: | Determining effects of experimental procedures on tissue involves careful observation of microscopic images. Human observation has potential for bias. In this project, I used three unsupervised machine learning methods to analyze H&E stained kidney tissue. |
| Methods: | I built an autoencoder for anomaly detection using Keras. The model includes an encoder that reduces dimensionality and a decoder that reconstructs the original input from the compressed representation. After training with H&E stained kidney tissue images, I calculated image reconstruction errors and set an error threshold for detecting anomalies.
To use the Anomalib library for anomaly localization, I split the images into training, testing, and validation sets. I created a Reverse Distillation model, instantiated an engine, and trained the model. This process generates a heatmap for an image, with anomalies highlighted. The Histomorphological Phenotype Learning framework begins with training a self-supervised model. Then, I used the model to create vector representations for each tile. Tiles are clustered using Leiden Clustering. This process outputs clustering configuration files as well as cluster assignments for individual tiles. |
| Results/Findings: | The autoencoder approach was ineffective. Reconstructed tissue images did not contain the necessary level of detail to be used for meaningful analysis.
The Anomalib library showed moderate success in identifying features of the kidney tissue. However, the generated heatmaps did not provide any insights. The Histomorphological Phenotype Learning (HPL) framework presented challenges due to technical issues. I am in the process of using the trained model to cluster images. |
| Conclusions/Discussion: | Each method had its strengths and limitations in detecting anomalies and identifying distinct features within the tissue images. |
| Translational/Human Health Impact: | Considering its ability to provide unbiased assessments and find patterns, the potential of unsupervised machine learning techniques for analyzing tissue images is significant. |
