Acute Lymphoblastic Leukemia (ALL) Detection Using Deep Learning Models from PBS Images

Local Chapter Irbid, Jordan Chapter

Coordinated byJordan ,

Rasha Awawdeh

Status: Completed

Project Duration: 02 Feb 2023 - 02 Mar 2023

Open Source resources available from this project

Github Repository

Project background.

Acute lymphocytic leukemia (ALL) is a type of cancer of the blood and bone marrow — the spongy tissue inside bones where blood cells are made.
The word “acute” in acute lymphocytic leukemia comes from the fact that the disease progresses rapidly and creates immature blood cells, rather than mature ones. Leukemic cells spread in the blood quickly and spread out to different parts of the body like spleen, liver, lymph nodes, brain, and nervous system. ALL is one of the top common type childhood cancer affecting Jordan patients and treatments have a good chance of curing it.

It is a new and exciting time for acute lymphoblastic leukemia (ALL). While nearly 50 years ago, only one in nine children with ALL survived with chemotherapy, nowadays nearly 90% of children have a chance of long-term survival. Adults with ALL, as well as the special category of adolescents and young adult (AYA) patients, are catching up with the new developments seen in children, but still their prognosis is much worse [1]. So that there is a huge need to raise awareness of the serious consequences and the importance of early diagnosis of ALL.

ALL is caused by a variety of genetic abnormalities, including mutations, chromosome translocations, and aneuploidy in genes involved in lymphoid cell development and cell cycle regulation. This work reviewed the medical records of children 1-18 years of age who were diagnosed with ALL and treated at King Hussein Cancer Center (KHCC) from January 2003 through December 2009. Disease characteristics and outcome were analysed. The results showed that over a 7-year period, 300 children with ALL were treated. One hundred and seventy-three (57.7%) were males and 127 (42.3%) were females. The median age at diagnosis was 5 years. One hundred and fifty-seven (52.3%) children were classified as low-risk, 118 (39.3%) were standard-risk and 25 (8.3%) were high-risk [3].

In order to implement automatic analysis for PBS images to detect ALL patients at early stages, we need to understand the main features that distinguish the ALL effect and variations over benign PBS images. According to the French-American-British (FAB) classification, acute lymphoblastic leukemia may be classified into 3 subgroups: ALL-L1 (small uniform cells), ALL-L2 (large varied cells) and ALL-L3 (large varied cells with vacuoles[4].

1. L1 – Around 25 to 30% of adult cases and 85% of childhood cases of ALL are of this subtype. In this type small cells are seen with:-
– Regular nuclear shape.
– Homogeneous chromatin.
– Small or absent nucleolus.
– Scanty cytoplasm.

2. L2 (large varied cells) – Around 70% of adult cases and 14% of childhood cases are of this type. The cells are large and or varied shapes with:-
– Irregular nuclear shape.
– Heterogeneous chromatin.
– Large nucleolus.

3. L3 (large varied cells with vacuoles) – This is a rarer subtype with only 1 to 2% cases. In this type the cells are large and uniform with vacuoles (bubble like features) in the cytoplasm overlying the nucleus.

The problem.

Acute leukemia is a life-threatening disease common in children and adults that can lead to death if left untreated. Acute lymphoblastic leukemia (ALL) spreads out in children’s bodies rapidly and takes their lives within a few weeks [2].

The definitive identification of acute lymphoblastic leukemia (ALL) needs, invasive, costly, and time-consuming diagnostic procedures. A critical step in the early separation of cancer cases from non-cancer cases is ALL diagnosis using peripheral blood smear (PBS) images [5]. The manual diagnosing method is completely reliant on professionally trained medical specialists and their experience. And this makes the examination fraught with problems, such as diagnostic mistakes. Many other factors contribute to misdiagnosis, such as poor peripheral blood smear (PBS) machine quality images or the non-specific nature of ALL signs and symptoms. This project investigates the opportunity to explore the best machine learning techniques to improve the ALL diagnosis by analyzing the best features of the disease. 

Project goals.

The project will explore the computer vision image filters until reaching the best filtered PBS image to extract the features. Multiple machine learning and deep learning models will be tested over the extracted features, and the best accuracy result will be adopted.   With a duration of 4-weeks, this project aims to:   1. Analyze the medical data images related to ALL. 2. Detect the best features related to ALL PBS images. 3. Investigate the best approaches to extracting the selected features in (2). 4. Apply multiple machine learning and deep learning models over the extracted features to find the model with the highest accuracy for classifying the PBS images. 5. Compare the results with previous works and analyze the output. 6. Generate a recommendation report

Project plan.

  • Week 1

    PBS Image preprocessing/Filtering

  • Week 2

    Extracting PBS image features for disease level

  • Week 3

    Apply pretrained DL algorithms over the extracted feature to classify PBS images

  • Week 4

    1- Apply pretrained models to extract features and ML models to classify the images and compare the results. 2- Generate final report and recommendation/ Evaluate Model accuracy

Learning outcomes.

1. Pre-processing the PBS image by applying some filters, such as: Active Contour.
2. Explore and detect the disease pattern features.
3. Implementing different deep learning models to classify PBS images, like ResNet152 or vgg16.
4. Test Convolutional Neural Network as a feature extractor with different ML classifiers and compare the results.
5. Provide a recommendation report on the best classifier for ALL PBS images.

Share project on: