leoll2/autoparking

Self parking with reinforcement learning

MedicalCNN

||

This project aims to perform abnormality classification in mammography by means of Convolutional Neural Networks. The dataset of interest is the CBIS DDSM dataset. The mammogram images feature two kinds of breast abnormalities: mass and calcification, which can be either benign or malignant.
The classification task consists in distinguishing between the four cases:

• Benign mass
• Malignant mass
• Benign calcification
• Malignant calcification

A subtask is to just distinguish masses from calcifications.

The full detailed report is available here.

Left: example of mass Right: example of calcification

---

Repo structure

All the Jupyter notebooks used for the experiments are collected in the scripts folder.

Specifically:

• Scratch_CNN_2_class: CNN built from scratch for the 2-categories classification task.
• Scratch_CNN_4_class: CNN built from scratch for the 4-categories classification task.
• Scratch_CNN_ben_mal: CNN build from scratch for benign-malignant classification.
• VGG16_2_class: VGG16 with feature-extraction and fine-tuning for the 2-categories classification task.
• VGG16_4_class: VGG16 with feature-extraction and fine-tuning for the 4-categories classification task.
• Baseline_Dual_CNN: Dual CNN model exploiting images of nearby healthy tissue too.
• Composite_4_class: Two parallel CNN models to decompose the 4-categories classification task.
• Baseline_Siamese: Siamese CNN exploiting images of nearby healthy tissue too.
• Ensemble_2_class: Ensemble of different CNN models for the 2-categories classification task.
• Ensemble_4_class: Ensemble of different CNN models for the 4-categories classification task.

Extra:

• LearningRate: Experiments tuning the learning rate for different optimizers.

You can download the dataset from Google Drive. All the scripts assume that the dataset zip file is located in the root of your Google Drive folder, but you can easily change it.

---

Experiments and results

I developed and tested many models for the 2-class and 4-class tasks.

The best model for the 2-class task obtained a 91.37% accuracy on the test set. The best model for the 4-class task obtained a 61.01% accuracy on the test set.

Comparing the results with those presented in many papers, the models achieved state-of-the-art accuracy [1][2][3].

[1] Neeraj Dhungel, Gustavo Carneiro, and Andrew P Bradley. “Automated mass
detection in mammograms using cascaded deep learning and random forests”.
In: 2015 international conference on digital image computing
[2] Dina A Ragab et al. “Breast cancer detection using deep convolutional neural
networks and support vector machines”. In: PeerJ 7 (2019)   
[3] Li Shen et al. “Deep learning to improve breast cancer detection on screening
mammography”. In: Scientific reports 9.1 (2019)  

See the report for full details.

---

Tools

The project was developed using the following technologies:

• Python: scripting language
• Keras: open-source library for experimentation with deep neural networks
• Google Colab: free cloud-based Jupyter notebook environment by Google

---

About the dataset

The dataset of interest is the CBIS DDSM (Curated Breast Imaging Subset of Digital Database for Screening Mammography), a collection of mammography images by Lee et al. It is an updated version of the original DDSM dataset, where all the images have been segmented and labeled.

Rebecca Sawyer Lee, Francisco Gimenez, Assaf Hoogi , Daniel Rubin  (2016). Curated Breast Imaging Subset of DDSM [Dataset]. The Cancer Imaging Archive. DOI: 10.7937/K9/TCIA.2016.7O02S9CY

---

Credits

The author (Leonardo Lai) designed and performed all the experiments listed in the project.

If you want to cite this work, please use the following:

@software{leonardo_lai_2021_4700130,
  author       = {Leonardo Lai},
  title        = {leoll2/MedicalCNN: v1.0},
  month        = apr,
  year         = 2021,
  publisher    = {Zenodo},
  version      = {1.0},
  doi          = {10.5281/zenodo.4700130},
  url          = {https://doi.org/10.5281/zenodo.4700130}
}

autoparking's Issues

allegro5 make problem

hello leoll2, I just install the allegro5 as you instructed, but when I run the make command ,there are some problems. before the make all dependencies have been install and update as shown in terminal.the problem log is below.by the way,my system is ubuntu 18.04 and python 3.7.5.

/Autoparking/allegro5/src/opengl/extensions.c:799:25: error: ‘ALLEGRO_OGL_EXT_API {aka struct ALLEGRO_OGL_EXT_API}’ has no member named ‘GetString’
glGetString = ext_api->GetString;

Spawning right next to the wall

If the vehicle spawns with one side perfectly adjacent to the wall, it has nowhere to go but forward or backwards. This results in a endless loop which, despite making sense, is kinda annoying to spectate and prevents other simulations from being performed.
Must be addressed asap!

i have issue while build the project

how to solve this i am working on this project

allegro5 马克

i have issue while build the project

how to solve this i am working on this project

Add pictures/textures

The environment would look fancier with additional colors, textures or pictures.

Cache needs a fix

Storing/restoring the matrix Q from cache doesn't seem to work properly, as the behavior of the network isn't quite the same before and after.
Needs investigation.

"Soft" collisions during the simulation

Sometimes the vehicle slightly goes out of the boundaries or touches obstacles.
This is not expected behavior, therefore has to be investigated.