Contents of folder:

This project is organized with the following directory structure:

Diagnosis_breast_cancer_MRI/

• This directory contains MRI scans used for breast cancer diagnosis.

MRI/

• This directory holds preprocessed MRI images in NIfTI format (.nii.gz) for analysis. Specific filenames suggest different acquisition sequences or orientations. This folder contains an axial MRI image of a breast. The image is obtained from a public dataset released by Duke University (https://sites.duke.edu/mazurowski/resources/breast-cancer-mri-dataset/)

  • T1_axial_slope1.nii.gz
  • T1_axial_slope2.nii.gz
  • T1_axial_01.nii.gz
  • T1_axial_02.nii.gz

Model/

• This directory stores the trained model for breast cancer diagnosis:
  • CNN_weights.npy - This file contains the weights for a Convolutional Neural Network (CNN).

Model Architecture The breast cancer detection model is a UNet architecture with the following specifications:

• Input shape: (512, 512, 3)
• Pooling size: (2, 2)
• Initial learning rate: 1e-5
• Depth: 6
• Base filters: 42
• Activation: Softmax
• L2 regularization: 1e-5
• Use of clinical data: True

Model Weights The pre-trained weights for this model are stored in CNN_weights.npy. These weights can be loaded into the model using the following code:

import numpy as np
from model_utils import FocalLoss, UNet_v0_2D_Classifier

model_weights = np.load('/path/to/your/model/CNN_weights.npy', allow_pickle=True)

model = UNet_v0_2D_Classifier(input_shape=(512, 512, 3), pool_size=(2, 2), initial_learning_rate=1e-5,
                             deconvolution=True, depth=6, n_base_filters=42,
                             activation_name="softmax", L2=1e-5, USE_CLINICAL=True)
model.set_weights(model_weights)

Scripts/

• This directory holds scripts for various tasks:
  • model_utils.py - This script contain utility functions for working with the model.
  • make_diagnosis_on_MRI.py - This script is used to make predictions on unseen MRI scans.
  • train_model.py - This script is used to train the CNN model from scratch or fine-tune existing weights.

Training_data/

• This directory holds training data used to train the model:

  • Data_Description.csv - The specific data here was generated at random and only serves the purpose of demonstrating the expected data format:

    • DE-ID: De-identified patient identifier. In this case just a number
    • Exam: Identifier for exam when image was taken
    • Scan_ID : Identifier of a specific scan: Constructed from DE-ID + Exam + side (in case of a saggital image)
    • Partition : Describes whether image is part of the training or validation set
    • Image : Points to location of the specific file containing the image of a scan. Name is constructed using DE-ID + Exam + Scan_ID + slice number of image.
    • BIRADS : From clinical information of exam
    • Pathology : From clinical information of exam
    • Age : From clinical information of exam
    • Family Hx : From clinical information of exam
    • Ethnicity and Race : From clinical information of exam

  /images/

• Description: This folder contains the necessary files for a new training session. It includes a spreadsheet with information on the training data and a folder with images to train the network. These images are examples taken from the same public dataset from Duke University. They are saved as PNG images for easier visualization. Note that due to the RGB color format, the images may appear unusual. The naming convention for the images is: Subject_Exam_Side_SliceNumber. The examples have fake names to serve as an example

  • SUBJECT2_EXAM1_l_14.png, etc

Sessions/

• This folder is used to store the output of each new training session. It saves the trained models, evaluation metrics, and any other relevant information generated during the training process.

Details of the pre-trained model

Training Data This network was trained on sagittal breast MRI taken at a tertiary cancer center between 2002-2014. The data included 38,005 exams (31,564 screening, 6,015 diagnostic, 426 unknown or N/A) from 12,329 patients. Counting each breast individually yielded 65,105 sagittal breast images with 2,690 malignant images. This data was randomly divided by patient into training, validation, and test sets (90/10 for training and test, and subsequently 90/10 for training and validation patients). This dataset included radiologist annotations/segmentation on the slice containing the largest (index) cancer for all 2,690 malignant breast images (termed the "index slice").

Data pre-processing and harmonization Preprocessing followed our previous work in segmentation. Briefly, pre- and post-contrast T1-weighted images were co-registered using NiftyReg, and dynamic contrast enhancement was summarized into images capturing initial contrast uptake (DC-in) and washout (DC-out), alongside the first post-contrast T1-weighted image (T1-post). These three channels were normalized by dividing by the 95th percentile of the pre-contrast T1-weighted image in each exam. To adjust for inter-channel differences, each channel was divided by its 95th percentile across the training set. This resulted in factors of 40 for T1-post, 0.3 for DCE-in, and 0.12 for DCE-out, applied to both sagittal and axial MRI. The sagittal MRI data had varying in-plane resolutions (0.4mm to 0.8mm). Low-resolution images were upsampled by a factor of two for harmonization. Axial images were resampled to match a sagittal in-plane resolution of 0.4mm and separated into left and right breasts. All images were cropped to 512x512 pixels, ensuring the breast was centered.

Training The model is trained using index slices from malignant images as positive examples. As negative examples, we selected the center slice and one randomly selected slice from benign images. All models were trained using a focal loss with alpha=5, using the ”Adam” optimizer with learning rate=1e-5. All models were trained for 100 epochs, and the model with the lowest validation loss was saved for evaluation. Unless specified, all models were trained with data augmentation, consisting of random rotation within 60 degrees, random shear of scale 0.1, random horizontal and vertical flips, and random intensity scaling in range 0.8-1.2, all implemented using the TensorFlow preprocessing ImageDataGenerator library.