Classifying histopathology slides of Lymphoma as malignant or benign using Convolutional Neural Network(CNN)
This project serves as a demonstration of how deep convolutional neural networks can achieve high accuracies in cancer histopathological image classification. The data used in this model was from Cross Cancer Institute, Edmonton, AB, curated by pathologist Dr.Gilbert Bigras. He digitized each of the 113 Lymphoma MYC IHC slides and labeled images as benign and malignant. The project was supervised by Dr.Nilanjan Ray.
Neural networks are a set of algorithms, modeled loosely after the human brain, that are designed to recognize patterns. They interpret sensory data through a kind of machine perception, labeling or clustering raw input. The patterns they recognize are numerical, contained in vectors, into which all real-world data, be it images, sound, text or time series, must be translated. (Skymind.ai)
The lymph system is a series of lymph nodes and vessels that move lymph fluid through the body. Lymph fluids contain infection-fighting white blood cells. Lymph nodes act as filters, capturing and destroying bacteria and viruses to prevent infection from spreading.
While the lymph system typically protects your body, lymph cells called lymphocytes can become cancerous. The names for cancers that occur in the lymph system are lymphomas.
Doctors classify more than 70 cancer types as lymphomas. Lymphomas can affect any portion of the lymphatic system, including:
- bone marrow
- thymus
- spleen
- tonsils
- lymph nodes
Doctors typically divide lymphomas into two categories:
- Hodgkin’s lymphoma
- non-Hodgkin’s lymphoma (NHL).
Cancer is the result of uncontrolled cell growth. The average lifespan of a cell is brief, and then the cell dies. In people with lymphoma, however, the cell thrives and spreads instead of dying. It’s unclear what causes lymphoma, but a number of risk factors are connected with these cancers.
A biopsy typically is taken if a doctor suspects lymphoma. This involves removing cells from an enlarged lymph node. A doctor known as a hematopathologist will examine the cells to determine if lymphoma cells are present and what cell type they are. If the hematopathologist detects lymphoma cells, further testing can identify how far the cancer has spread. These tests can include a chest X-ray, blood testing, or testing nearby lymph nodes or tissues. Imaging scans, such as a computed tomography (CT) or magnetic resonance imaging (MRI) scans may also identify additional tumors or enlarged lymph nodes.
The two major lymphoma types are Hodgkin’s lymphoma and non-Hodgkin’s lymphoma (NHL). A pathologist in the 1800s named Dr. Thomas Hodgkin identified the cells in what is now called Hodgkin’s lymphoma. Those with Hodgkin’s lymphoma have large cancerous cells called Reed-Sternberg (RS) cells. People with NHL don’t have these cells.
According to the Leukemia & Lymphoma Society (LLS), NHL is three times more common than Hodgkin’s lymphoma.Many lymphoma types fall under each category. Doctors call NHL types by the cells they affect, and if the cells are fast- or slow-growing. NHL forms in either the B-cells or T-cells of the immune system. According to LLS, most NHL types affect B-cells. Learn more about this type of lymphoma, who it affects, and where it occurs. Types include:
- B-cell lymphoma (85% NHL case)
Diffuse large B-cell lymphoma (DLBCL) is the most aggressive type of NHL. This fast-growing lymphoma comes from abnormal B cells in the blood. It can be cured if treated, but if left untreated, it can lead to death. The stage of DLBCL helps determine your prognosis. Read more about the stages and how this lymphoma is treated.
- T-cell lymphoma
T-cell lymphoma is not as common a B-cell lymphoma; only 15 percent of all NHL cases are this type. Several types of T-cell lymphoma exist. Learn more about each one, what they cause, and who is more likely to develop them.
- Burkitt’s lymphoma
Burkitt’s lymphoma is a rare type of NHL that is aggressive and most common in people with compromised immune systems. This type of lymphoma is most common in children in sub-Saharan Africa, but it does occur in other parts of the world. Learn more about this rare type of non-Hodgkin’s lymphoma.
- Follicular lymphoma
One in 5 lymphomas diagnosed in the United States are follicular lymphoma. This type of NHL, which starts in the white blood cells, is most common in older individuals. The average age of diagnosis is 60. This lymphoma is also slow-growing, so treatments begins with watchful waiting. Read more about this strategy.
- Primary mediastinal B cell lymphoma
This subtype of B-cell lymphoma accounts for almost 10 percent of DLBCL cases. It predominantly affects women in their 20s and 30s.
- Small lymphocytic lymphoma
Small lymphatic lymphoma (SLL) is a type of slow-growing lymphoma. The cancer cells of SLL are found mostly in the lymph nodes. SLL is identical to chronic lymphocytic leukemia (CLL), but with CLL, the majority of cancer cells are found in the blood and bone marrow.
Hodgkin’s lymphomas typically start in B-cells or immune system cells known as Reed-Sternberg (RS) cells. While the main cause of Hodgkin’s lymphoma is not known, certain risk factors can increase your chances of developing this type of cancer. Learn what these risk factors are. Hodgkin’s lymphoma types include:
- Lymphocyte-depleted Hodgkin’s disease
This rare, aggressive type of lymphoma occurs in about 1 percent of lymphoma cases, and it’s most commonly diagnosed in individuals in their 30s. In diagnostic tests, doctors will see normal lymphocytes with an abundance of RS cells. Patients with a compromised immune system, such as those with HIV, are more likely to be diagnosed with this type of lymphoma.
- Lymphocyte-rich Hodgkin’s disease
This type of lymphoma is more common in men, and it accounts for about 5 percent of Hodgkin’s lymphoma cases. Lymphocyte-rich Hodgkin’s disease is typically diagnosed at an early stage, and both lymphocytes and RS cells are present in diagnostic tests.
- Mixed cellularity Hodgkin’s lymphoma
Like with lymphocyte-rich Hodgkin’s disease, mixed cellularity Hodgkin’s lymphoma contains both lymphocytes and RS cells. It’s more common — almost a quarter of Hodgkin’s lymphoma cases are this type — and it’s more prevalent in older adult men.
- Nodular lymphocyte-predominant Hodgkin’s disease
Nodular lymphocyte-predominant Hodgkin’s disease (NLPHL) type of Hodgkin’s lymphoma occurs in about 5 percent of lymphoma patients, and it’s characterized by an absence of RS cells. NLPHL is most common in people between the ages of 30 and 50, and it’s more common in males. Rarely, NLPHL can progress or transform into a type of aggressive NHL.
- Nodular sclerosis Hodgkin’s lymphoma
This common type of lymphoma occurs in 70 percent of Hodgkin’s cases, and it’s more common in young adults than any other group. This type of lymphoma occurs in lymph nodes that contain scar tissue, or sclerosis.
A total of 113 lymphomas (19 Burkitt’s lymphoma(BL), 77 Diffuse large B-cell lymphoma(DLBCL), 6 intermediate between BL and DLBCL, and 11 unclassified aggressive B-cell lymphomas) diagnosed between 2010 and 2015 with known MYC status were selected. MYC IHC stains were produced in 2014 and 2015. All specimens had been fixed in formalin (37% formaldehyde in aqueous solution) and embedded in paraffin. Although a small group of cases (B20%) was studied retrospectively, all cases in this analysis were stained utilizing freshly cut thin sections.
The dataset consisted of 4 types of 113 whole mount slide images (2560x1920) of Lymphoma specimens scanned at 20x. From each type of master image, around 42,750 sequential non-overlapping child patches of size 100 x 100 were extracted (roughly 23,275 negative and 19,475 positive).
Each type of mother image file name is of the format: type_patientID_class.tif
| Example | Example | Image Type | Type serial |
|---|---|---|---|
| 2_neg.tif | 7_pos.tif | MYC IHC | 0 |
| dab_2_neg.tif | dab_7_pos.tif | DAB (MYC signal) | 1 |
| dist_2_neg.tif | dist_7_pos.tif | Distance map of positive nuclei | 2 |
| hem_2_neg.tif | hem_7_pos.tif | Hematoxylin (blue counterstrain) | 3 |
| Benign(-) | Malignant(+) |
|---|---|
| MYC IHC | MYC IHC |
 |
 |
| DAB(MYC signal) | DAB(MYC signal) |
 |
 |
| Distance map of positive nuclei | Distance map of positive nuclei |
 |
 |
| Hematoxylin(blue counterstrain) | Hematoxylin(blue counterstrain) |
 |
 |
Two different model architectures are compared on the basis of performance on train-test dataset generated with 113 type-0 (MYC IHC) master images. The best one, VGG19 is selected.
| Benign(-) | Malignant(+) |
|---|---|
| MYC IHC | MYC IHC |
|
|
More details: Model Architecture Comparison
The whole process is divided into 10 TestRuns to assure performance consistency of the models. 10 testruns are performed for each model where they are trained and evaluated on different train-test dataset generated from shuffled raw set of master images.
Size of trainset is 90x4 and testset is 23x4 out of 113x4 master images. 113 mother images of a type is shuffled, then randomly 90 images of that type along with other 3 types holding same patientID are selected for trainset(90x4).
For each of the 10 TestRuns, there are 3 different versions of trained model on the basis of combination of 4 types of 113 master images.
| Training Version | Model Name | Image Type | Type serial | Sample | Training History (TestRun1) |
|---|---|---|---|---|---|
| 1 | VGGNet19 | MYC IHC | 0 | |
 |
| 2 | VGGNet19v2 | MYC IHC DAB (MYC signal) Distance map of positive nuclei Hematoxylin (blue counterstrain) |
0 1 2 3 |
|
 |
| 3 | VGGNet19v4 | MYC IHC DAB (MYC signal) Hematoxylin (blue counterstrain) |
0 1 3 |
|
 |
GPU on google colab: Tesla P100
- version 1:
- 1.3 sec per image(full size)
- version 2:
- 7 sec per image(full size)
- version 3:
- 4.5 sec per image(full size)
Test result of each of the 3 versions of model on each of the 10 testruns with testset size of 23 full size image.
- Supervisors:
- Dr.Nilanjan Ray, Associate Professor, Department of Computing Science, University of Alberta, Canada
- Dr.Gilbert Bigras, Associate Professor, Department of Laboratory Medicine & Pathology, University of Alberta, Canada
- Dataset: https://bit.ly/2MdWSzp
- Citation:
- Reference: https://www.healthline.com/health/lymphoma#diagnosis
Please send me your feedback at sandipsahajoy@ualberta.ca