Home
Predicting Flight Delays Exploring Catastrophic Forgetting for Question Answering Domain Adaptation in Text-to-Text Transformers Effect of Political Ideology on Consumer Behavior Pulmonary Nodule Segmentation Poverty and COVID-19 Barriers to entry in Open Source Software Certifications and Discounts Interactive Wheel Based Image Recoloring GUI Coreless Promoters in Arabadopsis DNA Health Portal - Built for All

Pulmonary Nodule Segmentation

The Purpose

Our purpose was given the CT scans of a lung we want to predict the location(s) of the nodules in the lung. Currently, identification and classification of lung nodules are done by experienced annotators. We aim to build a model that can segment a lung CT scan and predict the existence of a nodule.

The dataset we are using is from a public dataset called Lung Nodule Analaysis 2016 (LUNA-16). It has a total of 888 CT scans. There are a total of 1186 nodules for 601 of those CT scans. The annotations file contatins the x,y,z coordinates of each nodule with a size >= 3mm and its corresponding CT scan. The data files take up approximately 115 GB of data.

The Quest

  • Locate the nodules to use as input
    • Using the annotated data, we know for each CT Scan where the nodules in that image are.
    • Identify the nodules by matching the annotated locations given by the experts to the images.
    • A subset of these images (training set) will be processed so that a nodule mask will be generated for each nodule.
    • The nodule masks will be inputted into the model.
  • Classify each pixel in image (Semantic Segmentation)
    • Input CT scans into the model so the model can be trained.
    • Models will classify each pixel in the image and label it as a nodule or not.
  • Once model is built we can use it to predict the location of lung nodules in other CT scans of lungs (testing set).

The Solution

Semantic Segmentation

Semantic Segmentation is the process of taking an image and labeling each pixel as a type of object. In our case, the object will either be a nodule or not a nodule

Understanding our Dataset

Nodules labeled by expert annotators on the lung CT scans
Nodules distribution by size and frequency

Preprocessing Images

To identify lung tissue create a lung mask (Since lungs are full of air it is less dense in CT scans)
To reduce size and noise of data identify a "region of interest" (i.e. the lungs) for the model to focus on

Modeling and Training

We used a "U-Net" architecture for our Convolutional Neural Network (CNN) which consists of 2 branches, contracting and expanding.

Semantic Segmentation is the process of taking an image and labeling each pixel as a type of object. In our case, the object will either be a nodule or not a nodule
The base model parameters we built using the U-Net architecture were as follows:

We used tensorflow_cloud api to train the model in Google AI Platform.

The Impact

Performance of the Base Model (the training data had a dice coefficent of 0.799 and the validation data had a dice coefficent of 0.690.
The base model parameters we built using the U-Net architecture were as follows:
Here a subset of predicted images next to annotated images: