BREAST CANCER PREDICTION FROM MAMMOGRAMS

Anveshak Rathore	Akshay Kolli	Prathyusha Reddy Nandikonda	Lalitha Spoorthy Dantuluri
Miner School of Computer & Information Sciences, Kennedy College of Sciences, University of Massachusetts, Lowell	Miner School of Computer & Information Sciences, Kennedy College of Sciences, University of Massachusetts, Lowell	Miner School of Computer & Information Sciences, Kennedy College of Sciences, University of Massachusetts, Lowell Lowell	Miner School of Computer & Information Sciences, Kennedy College of Sciences, University of Massachusetts, Lowell
anveshak_rathore@student.uml.edu	akshay_kolli@student.uml.edu	prathyushareddy_nandikonda@student.uml.edu	LalithaSpoorthy_Dantuluri@student.uml.edu

Abstract:

Early detection of diseases has recently become an important concern in medical research due to rapid growth in the population. Breast cancer is the second most fatal of the cancers that have already been identified. An automatic disease detection system assists healthcare professionals in disease diagnosis, provides reliable, efficient, and rapid action, and reduces the risk of death.

In this study, we are proposing a breast cancer prediction model using mammograms using a convolutional autoencoder. The autoencoder is trained on exclusively non- cancer images, such that it perfectly reconstructs a non- cancer image with minimal loss. This way, when the model encounters a cancer positive image, it will give high loss on reconstruction which can be interpreted as a possible positive case. We aim to create a tool that will help the medical community in the pre-screening phase, to minimize human error and effort in recognizing possible positive cases. We analyze the outcomes of the stated approaches using measures such as Accuracy, F1 score, and Confusion Matrix.

1. INTRODUCTION:

Breast cancer is the largest cause of cancer-related deaths among women around the world. 1in 8 women is affected by breast cancer in their lifetime as per American Cancer Society Survey Report in 2013. Early detection and precise diagnosis are crucial for improving the health of patients and decreasing mortality rates. Mammography is the most widely used imaging modality for screening breast cancer. However, mammography interpretation might be difficult.

In this project, we propose a breast cancer prediction model based on mammograms that use a convolutional autoencoder. An autoencoder is a neural network design that can be trained to compress input data into a lower- dimensional latent space and then recover the original data from this space. The autoencoder is trained solely on non-cancer images, allowing it to rebuild a non- cancer image with minimum loss. When the model encounters a cancer positive image, it will return a large loss on reconstruction, indicating a likely positive case.

Class imbalance is a major problem in disease prediction in machine learning. It refers to a situation in which the number of examples in one class is much less than the number of instances in another class. Class imbalance is a common issue in the field of disease classification using medical imaging. Images from affected patients are significantly harder to come by in the case of rare medical disorders than images from non- affected patients, resulting in an undesired class imbalance. For beneficial results, the class distribution must be balanced while training a model. This problem serves as the foundation for our approach to using an autoencoder for mammogram classification.

2. RELATED WORK:

1. Aminul Huq, Md Tanzim Reza, Shahriar Hossain, Shakib Mahmud Dipto, AnoMalNet: Outlier Detection based Malaria Cell Image Classification Method Leveraging Deep Autoencoder. International Journal of Reconfigurable and Embedded Systems (IJRES), Vol. 99, No. 1, Month 2099, pp. 1∼1x.
2. Kim, H.E., Cosa-Linan, A., Santhanam, N. et al. Transfer learning for medical image classification: a literature review. BMC Med Imaging 22, 69 (2022).
3. S. Guan and M. Loew, "Breast Cancer Detection Using Transfer Learning in Convolutional Neural Networks," 2017 IEEE Applied Imagery Pattern Recognition Workshop (AIPR), Washington, DC, USA, 2017, pp. 1-8.
4. Q. A. Al-Haija and A. Adebanjo, "Breast Cancer Diagnosis in Histopathological Images Using ResNet-50 Convolutional Neural Network," 2020 IEEE International IOT, Electronics and Mechatronics Conference (IEMTRONICS), Vancouver, BC, Canada, 2020, pp. 1-7.
5. S. Kumar, A. Narang, M. Parihar and V. Sawant, "Breast Tumour Classification using MobileNetV2," 2021 2nd Global Conference for Advancement in Technology (GCAT), Bangalore, India, 2021, pp. 1-6.

3. METHDOLOGY:

Dataset Description:

We are using the RSNA Screening Mammography Breast Cancer detection Dataset. The Dataset consists of 4 images for each person with a whole size of 54707 images. In this Dataset we have 1158 of Positive samples and 53549 are of negative samples.

Cancerous images	Non-cancerous images

Dataset Preprocessing:

Re-organize: The original data is organized based on the patient. This is reorganized into images with and without cancer.

Resize: The image resolution is lowered from 5000 * 5000 to 500 * 500.

Reformat: The mammograms in the original dataset are in dicom format. These are converted to PNG format.

Fig 1: Original Vs Reconstructed Images

Approach:

Autoencoder

A custom encoder with the decoder will be trained using mammograms free of cancer images. The proposed approach is based on the intuition that, during testing, this trained auto-encoder will achieve a loss score for mammograms free of cancer. However, in the case of mammograms with cancer, the model will output a significantly higher loss value.

With the help of simple statistics, a cut-off point can be established to label unknown mammograms as the one with cancer or cancer free. An unknown mammogram is determined as an outlier if the loss value of the unknown image which we get after passing through the model is more than the mean plus three times of the standard deviation of train loss.

Fig 2: Auto Encoder- Training with Cancer free Mammograms

Fig 3: Auto-encoder: Testing with Cancer free Mammograms

Fig 4: Auto Encoder- Testing with cancer positive Mammogram

Transfer Learning

Transfer learning is a machine learning technique in which a model that has been trained on one task is re-purposed for another related task. The use of a pre-trained convolutional neural network (CNN) as a feature extractor is a common approach for transfer learning in breast cancer prediction. Another technique is to train a pre-trained CNN to predict breast cancer. Fine- tuning entails retraining the last couple of layers of a pre- trained convolutional neural network (CNN) using a smaller dataset of mammograms. This enables the network to customize features acquired from a bigger dataset to the specific task of breast cancer prediction.

Fine-tuning often involves freezing the weights of the previously learned CNN layers and just training the weights of the last few layers connected to a new classifier. To avoid overfitting, the learning rate is often set to a lower value than during pre-training. The number of epochs used for fine-tuning is often less than that used for pre-training.

Fig 5: Fine Tuning Architecture

• MobilenetV2: MobileNetV2 is a convolutional neural network with fewer operations than other architectures, allowing for faster model training. Because of its small size, high computing speed, and competitive performance, MobileNetV2 is also suited for mobile devices.

Fig 6: Architecture of MobileNetV2

• ResNet50: CNNs have at least one Convolution layer, wherein instead of matrix multiplication, a convolution operation is performed on the input matrix in order to learn distinct low-level and high-level features of the image. Deep CNNs are able to learn more features by increasing the depth of the network. However, increasing the depth of the network results in problems of vanishing gradients and degradation. Residual neural networks (ResNet) address these challenges by introducing a "Residual block", which features a "skip connection", that adds the output from the previous layer to the layer ahead as illustrated in Fig. 6. If x and F(x) below do not have the same dimension, x is multiplied by a linear projection W to equalize the dimensions of the short-cut connection and the output layer.

Fig 7: Residual Network building block

• VGG16: A CNN design known as the VGG16 is made up of a stack of convolutional layers with tiny 3x3 filters, max-pooling layers, and fully linked layers at the very end. There are 13 convolutional layers in all, the first two of which have 64 filters each, and the following 11 of which have 128 filters. The feature maps created by the convolutional layers are down sampled using the max-pooling layers, which keeps all of the crucial characteristics while lowering their spatial dimensions. With the features taken from the convolutional layers, the fully connected layers at the end of the VGG16 architecture are used for classification to get a class prediction. A probability distribution over the various classes is output by the softmax layer, which is the network's last layer.

Fig 10: VGG16-CNN Model

4. RESULTS:

We applied the trained models to the images in the test dataset to assess the outcome, i.e. whether a person has breast cancer or not. To determine how accurate the training models are, the outcome prediction was compared to the corresponding target feature in the testing set. To compare the models, we use a few model evaluation criteria from the scikit-learn Python module.

Autoencoder Accuracy: 99.218%

Transfer Learning:

Train images = 49235 images in 2 classes Validation images = 5471 images in 2 classes Input image size = 500x500x3

Validation Accuracies:

• VGG16: 98.92%
• Mobilenet V2: 97.92%
• ResNet50: 97.22%

Final autoencoder MSE loss = 0.0011

5. REFERENCES:

1. Dataset: RSNA Screening Mammography Breast Cancer Detection | Kaggle https://www.kaggle.com/competitions/rsna-breast-cancer-detection/data
2. Aminul Huq, Md Tanzim Reza, Shahriar Hossain, Shakib Mahmud Dipto, AnoMalNet: Outlier Detection based Malaria Cell Image Classification Method Leveraging Deep Autoencoder. International Journal of Reconfigurable and Embedded Systems (IJRES), Vol. 99, No. 1, Month 2099, pp. 1∼1x. https://arxiv.org/pdf/2303.05789.pdf
3. Kim, H.E., Cosa-Linan, A., Santhanam, N. et al. Transfer learning for medical image classification: a literature review. BMC Med Imaging 22, 69 (2022). https://link.springer.com/article/10.1186/s12880-022-00793-7#citeas
4. S. Guan and M. Loew, "Breast Cancer Detection Using Transfer Learning in Convolutional Neural Networks," 2017 IEEE Applied Imagery Pattern Recognition Workshop (AIPR), Washington, DC, USA, 2017, pp. 1-8. https://ieeexplore.ieee.org/abstract/document/8457948
5. Q. A. Al-Haija and A. Adebanjo, "Breast Cancer Diagnosis in Histopathological Images Using ResNet-50 Convolutional Neural Network," 2020 IEEE International IOT, Electronics and Mechatronics Conference (IEMTRONICS), Vancouver, BC, Canada, 2020, pp. 1-7. https://ieeexplore.ieee.org/abstract/document/9216455
6. S. Kumar, A. Narang, M. Parihar and V. Sawant, "Breast Tumour Classification using MobileNetV2," 2021 2nd Global Conference for Advancement in Technology (GCAT), Bangalore, India, 2021, pp. 1-6. https://ieeexplore.ieee.org/abstract/document/9587501
7. https://www.sciencedirect.com/science/article/abs/pii/B9780128181461000040