sara-i / CUBIG_Covid

0 stars 0 forks source link

training model #3

Open sara-i opened 2 years ago

sara-i commented 2 years ago

For น้องเก็ท & จีจี้ RTCU

sara-i commented 2 years ago

Model construction

โจทย์ :

  1. ทำนาย COVID positive/ negative
  2. สร้าง GradCam เพื่อ visualize บริเวณที่ทำให้คิดว่าเป็น covid positive
  3. Display bounding box ซ้อนทับกับ GradCam เพื่อให้มั่นใจว่า Gradcam เราทำนายถูก

Dataset:

For model training of COVID diagnosis COVID-19 Radiography Database

For GradCam model

For bounding box visualization and test set SIIM-FISABIO-RSNA COVID-19 Detection

sara-i commented 2 years ago

จาก SIIM kaggle dataset

Total = 6054 cases

  1. Pneumonia = 1676 cases
  2. Typical Appearance = 2855 cases
  3. Indeterminate Appearance = 1049 cases
  4. Atypical = 474 cases

มี bounding box = 4294/6054 cases

  1. Pneumonia = 0 cases
  2. Typical Appearance = 2854 cases
  3. Indeterminate Appearance = 1049 cases
  4. Atypical = 391 cases

สรุปว่าเอาเอาแบบ weight bounding boxes

typ 66/2854 Indet 24/1049 Atyp 10/391

sara-i commented 2 years ago

การสร้าง Model 1: COVID-19 diagnostic prediction

1. Dataset: COVID-19 Radiography Database (Normal = 10192, COVID = 3616) ไม่สามารถเอา dataset มาทั้งหมดได้เนื่องจากข้อจำกัดของ computer -> Normal = 2000, COVID = 2000

Split data into train-val-test : 70-20-10

2. Model: Densenet121

3. Weight: ImageNet

/////////////////////////// จบ ///////////////////////////

sara-i commented 2 years ago

การสร้าง Model 2: GradCam

1. Dataset: ต่อเนื่องจาก Model 1

2.Concept: คร่าวๆคือ จะเอา gradient ที่ได้บนภาพมา display

/////////////////////////// จบ ///////////////////////////

sara-i commented 2 years ago

การสร้าง Model 3: เอาทุกอย่างมา apply เข้าด้วยกัน

คือ ใส่ Dataset ใหม่เข้าไป สามารถทำนายได้ว่าเป็น COVID + ซ้อนทับด้วยภาพ GradCam + ซ้อนทับด้วยภาพ bounding box

1. Dataset: SIIM-FISABIO-RSNA COVID-19 Detection (Total = 4294 cases)

/////////////////////////// จบ ///////////////////////////