Chromatin is a structure present within the cell nucleus. The reason
behind studying chromatin structure is to analyze both efficacy and risk of cancer
drugs. When a section of DNA is being expressed or repaired, chromatin assumes
to be a more spread-out state.
Our research goal is to accurately identify the center position of chromatin structure
using Deep Learning models. we used an object detection model called Faster-RCNN
model to identify chromatin spirals, and subsequently derived spiral centers.While our results from using the FRCNN model were satisfactory, we noticed that
there were still some small pixel value differences in between the centers predicted
by the model and the actual centers (NOSPP centers). To bridge this gap and minimize errors, we proposed a solution: predicting the errors using Sequential
model, and then adjusting the coordinates accordingly.
In our current proposal, we aim to modify these centers (point coordinates) by
incorporating error adjustments. To achieve this, we utilized a Deep Learning
approach called Long Short-Term Memory (LSTM) model to predict errors
associated with the chromatin patterns.
Initially, we trained a deep learning model to learn knowledge of the error patterns
and during testing, if we pass a spiral the model must detect any errors. The
predicted errors are again passed back to adjust the predicted center point
coordinates by FRCNN model. Our plan is to update the predicted center
coordinates based on the predicted errors.