SIIM-ISIC-Melanoma
🔬 SIIM-ISIC-Melanoma-Classification
📝 Description
Skin cancer is the most prevalent type of cancer. Melanoma, specifically, is responsible for 75% of skin cancer deaths, despite being the least common skin cancer. The American Cancer Society estimates over 100,000 new melanoma cases will be diagnosed in 2020. It’s also expected that almost 7,000 people will die from the disease. As with other cancers, early and accurate detection—potentially aided by data science—can make treatment more effective.
Currently, dermatologists evaluate every one of a patient’s moles to identify outlier lesions or “ugly ducklings” that are most likely to be melanoma. Existing AI approaches have not adequately considered this clinical frame of reference. Dermatologists could enhance their diagnostic accuracy if detection algorithms take into account “contextual” images within the same patient to determine which images represent a melanoma. If successful, classifiers would be more accurate and could better support dermatological clinic work.
As the leading healthcare organization for informatics in medical imaging, the Society for Imaging Informatics in Medicine (SIIM)’s mission is to advance medical imaging informatics through education, research, and innovation in a multi-disciplinary community. SIIM is joined by the International Skin Imaging Collaboration (ISIC), an international effort to improve melanoma diagnosis. The ISIC Archive contains the largest publicly available collection of quality-controlled dermoscopic images of skin lesions.
In this competition, you’ll identify melanoma in images of skin lesions. In particular, you’ll use images within the same patient and determine which are likely to represent a melanoma. Using patient-level contextual information may help the development of image analysis tools, which could better support clinical dermatologists.
Melanoma is a deadly disease, but if caught early, most melanomas can be cured with minor surgery. Image analysis tools that automate the diagnosis of melanoma will improve dermatologists’ diagnostic accuracy. Better detection of melanoma has the opportunity to positively impact millions of people.
🔍 Evaluation
Submissions are evaluated on area under the ROC curve between the predicted probability and the observed target.
What is AUC - ROC Curve?
AUC - ROC curve is a performance measurement for classification problem at various thresholds settings. ROC is a probability curve and AUC represents degree or measure of separability. It tells how much model is capable of distinguishing between classes. Higher the AUC, better the model is at predicting 0s as 0s and 1s as 1s. By analogy, Higher the AUC, better the model is at distinguishing between patients with disease and no disease.
The ROC curve is plotted with TPR against the FPR where TPR is on y-axis and FPR is on the x-axis.
▶ Here is the link to a great video on roc-auc
📝 More About Melanoma
- The most serious type of skin cancer.
- Melanoma occurs when the pigment-producing cells that give colour to the skin become cancerous.
- Symptoms might include a new, unusual growth or a change in an existing mole. Melanomas can occur anywhere on the body.
- Treatment may involve surgery, radiation, medication or in some cases, chemotherapy.
✔ ABCDE Rule
Use the “ABCDE rule” to look for some of the common signs of melanoma, one of the deadliest forms of skin cancer:
✔ Ugly Duckling
The Ugly Duckling is another warning sign of melanoma. This recognition strategy is based on the concept that most normal moles on your body resemble one another, while melanomas stand out like ugly ducklings in comparison. This highlights the importance of not just checking for irregularities, but also comparing any suspicious spot to surrounding moles to determine whether it looks different from its neighbours. These ugly duckling lesions or outlier lesions can be larger, smaller, lighter or darker, compared to surrounding moles. Also, isolated lesions without any surrounding moles for comparison are considered ugly ducklings.
Melanomas commonly appear on the legs of women, and the number one place they develop on men is the trunk.
Many other factors also play a role in increasing the risk for melanoma, including genetics (family history), skin type or colour, hair colour, freckling and number of moles on the body.
✔ These factors increase your melanoma risk
- Unprotected or excessive UV exposure from the sun or indoor tanning.
- Weakened immune system due to a medical condition or medications.
- Many moles: The more moles you have on your body, the higher your risk for melanoma. Also, having large moles (larger than a tip of a pencil eraser), or any atypical moles, increases the risk for melanoma.
- Fair skin: Melanoma occurs more frequently in people with fair skin, light eyes and light or red hair.
- Skin cancer history: People who have already had melanoma or nonmelanoma skin cancers run a greater risk of developing melanoma in the future.
- Genetics: Melanoma can run in families – one in every 10 patients has a family member who also has had the disease.
✔ Melanoma Hair Remove
There are many images with body hair covering the lesion so, hair remove operation can be useful for model focus on lesion part. Method for hair remove using CV2
def hair_remove(image):
# convert image to grayScale
grayScale = cv2.cvtColor(image, cv2.COLOR_RGB2GRAY)
# kernel for morphologyEx
kernel = cv2.getStructuringElement(1,(17,17))
# apply MORPH_BLACKHAT to grayScale image
blackhat = cv2.morphologyEx(grayScale, cv2.MORPH_BLACKHAT, kernel)
# apply thresholding to blackhat
_,threshold = cv2.threshold(blackhat,10,255,cv2.THRESH_BINARY)
# inpaint with original image and threshold image
final_image = cv2.inpaint(image,threshold,1,cv2.INPAINT_TELEA)
return final_image
- this technique works on black-Hair only (Not white hair)
Example of body hair remove with this method:
My Model performace and Outcomes
Single model performace 2020 data [M]
| No | model | Image-Size | CV-Score | LB-Score | External Data | Details | parameters |
|---|---|---|---|---|---|---|---|
| 1 | EfficientNetB6 (noisy-student) |
512x512 | 0.913 | 0.9405 | ❌ | Fold-1 max_auc=0.90 Fold-2 max_auc=0.90 Fold-3 max_auc=0.88 |
Fold=3 epochs=15 TTA=15 INC2019 = [0,0,0] INC2018 = [0,0,0] |
| 2 | EfficientNetB6 (noisy-student) |
384x384 | 0.913 | 0.9389 | ❌ | Fold-1 max_auc=0.92 Fold-2 max_auc=0.90 Fold-3 max_auc=0.90 Fold-4 max_auc=0.88 Fold-5 max_auc=0.89 |
Fold=5 epochs=15 TTA=15 INC2019 = [0,0,0,0,0] INC2018 = [0,0,0,0,0] |
| 3 | EfficientNetB6 (noisy-student) |
256x256 | 0.908 | - | ❌ | cosine_schedule Focal_loss Fold-1 max_auc=0.918 Fold-2 max_auc=0.910 Fold-3 max_auc=0.900 |
Fold=3 epochs=20 TTA=15 INC2019 = [0,0,0] INC2018 = [0,0,0] |
| 4 | EfficientNetB6 (imagenet) |
192x192 | 0.904 | 0.9132 | ❌ | Fold-1 max_auc=0.907 Fold-2 max_auc=0.912 Fold-3 max_auc=0.895 |
Fold=3 epochs=20 TTA=20 INC2019 = [0,0,0] INC2018 = [0,0,0] |
| 5 | EfficientNetB6 (noisy-student) |
128x128 | 0.895 | 0.9253 | ❌ | Fold-1 max_auc=0.890 Fold-2 max_auc=0.905 Fold-3 max_auc=0.890 |
Fold=3 epochs=20 TTA=20 INC2019 = [0,0,0] INC2018 = [0,0,0] |
| 6 | EfficientNetB5 (noisy-student) |
512x512 | 0.919 | - | ❌ | cosine_schedule Focal loss Fold-1 max_auc=0.921 Fold-2 max_auc=0.930 Fold-3 max_auc=0.906 |
Fold=5 epochs=15 TTA=15 INC2019 = [0,0,0] INC2018 = [0,0,0] |
| 7 | EfficientNetB5 (noisy-student) |
384x384 | 0.918 | - | ❌ | cosine_schedule Focal_loss Fold-1 max_auc=0.924 Fold-2 max_auc=0.923 Fold-3 max_auc=0.909 |
Fold=5 epochs=15 TTA=15 INC2019 = [0,0,0,0,0] INC2018 = [0,0,0,0,0] |
| 8 | EfficientNetB5 (imagenet) |
256x256 | 0.909 | 0.9219 | ❌ | cosine_schedule Focal_loss Fold-1 max_auc=0.915 Fold-2 max_auc=0.914 Fold-3 max_auc=0.910 |
Fold=3 epochs=20 TTA=20 INC2019 = [0,0,0] INC2018 = [0,0,0] |
| 9 | EfficientNetB5 (noisy-student) |
192x192 | 0.908 | 0.9201 | ❌ | cosine_schedule Focal_loss Fold-1 max_auc=0.928 Fold-2 max_auc=0.919 Fold-3 max_auc=0.908 Fold-4 max_auc=0.885 Fold-5 max_auc=0.903 |
Fold=5 epochs=15 TTA=15 INC2019 = [0,0,0,0,0] INC2018 = [0,0,0,0,0] |
| 10 | EfficientNetB5 (noisy-student) |
128x128 | 0.908 | - | ❌ | lr_callback Focal_loss Fold-1 max_auc=0.92 Fold-2 max_auc=0.903 Fold-3 max_auc=0.915 Fold-4 max_auc=0.893 Fold-5 max_auc=0.907 |
Fold=5 epochs=15 TTA=15 INC2019 = [0,0,0,0,0] INC2018 = [0,0,0,0,0] |
| 11 | EfficientNetB4 (noisy-student) |
512x512 | 0.913 | - | ❌ | cosine_schedule Focal_loss Fold-1 max_auc=0.920 Fold-2 max_auc=0.918 Fold-3 max_auc=0.904 |
Fold=5 epochs=20 TTA=15 INC2019 = [0,0,0] INC2018 = [0,0,0] |
| 12 | EfficientNetB4 (noisy-student) |
384x384 | 0.920 | 0.9327 | ❌ | cosine_schedule Focal loss Fold-1 max_auc=0.921 Fold-2 max_auc=0.927 Fold-3 max_auc=0.929 Fold-4 max_auc=0.899 Fold-5 max_auc=0.924 |
Fold=5 epochs=15 TTA=15 INC2019 = [0,0,0,0,0] INC2018 = [0,0,0,0,0] |
| 13 | EfficientNetB4 (noisy-student) |
256x256 | 0.918 | - | ❌ | Lr_schedule Focal loss Fold-1 max_auc=0.920 Fold-2 max_auc=0.930 Fold-3 max_auc=0.935 Fold-4 max_auc=0.891 Fold-5 max_auc=0.917 |
Fold=5 epochs=15 TTA=15 INC2019 = [0,0,0,0,0] INC2018 = [0,0,0,0,0] |
| 14 | EfficientNetB4 (noisy-student) |
192x192 | 0.915 | - | ❌ | lr_schedule Focal loss Fold-1 max_auc=0.928 Fold-2 max_auc=0.913 Fold-3 max_auc=0.925 Fold-4 max_auc=0.896 Fold-5 max_auc=0.916 |
Fold=5 epochs=15 TTA=15 INC2019 = [0,0,0,0,0] INC2018 = [0,0,0,0,0] |
| 15 | EfficientNetB4 (noisy-student) |
128x128 | 0.907 | - | ❌ | Lr_schedule Focal loss Fold-1 max_auc=0.918 Fold-2 max_auc=0.912 Fold-3 max_auc=0.911 Fold-4 max_auc=0.893 Fold-5 max_auc=0.904 |
Fold=5 epochs=15 TTA=15 INC2019 = [0,0,0,0,0] INC2018 = [0,0,0,0,0] |
Single model performace 2020 + 2018 data [R]
| No | model | Image-Size | CV-Score | LB-Score | External Data | Details | parameters |
|---|---|---|---|---|---|---|---|
| 1 | EfficientNetB6 (imagenet) |
512x512 | 0.930 | - | ✔️ | cosine_schedule Focal loss Fold-1 max_auc=0.925 Fold-2 max_auc=0.931 Fold-3 max_auc=0.936 |
Fold=3 epochs=20 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
| 2 | EfficientNetB6 (imagenett) |
384x384 | 0.917 | 0.9417 | ✔️ | Upsampling M1,M3,M4 cosine_schedule Focal loss Fold-1 max_auc=0.910 Fold-2 max_auc=0.925 Fold-3 max_auc=0.918 |
Fold=3 epochs=20 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
| 3 | EfficientNetB6 (imagenet) |
256x256 | 0.913 | - | ✔️ | cosine_schedule Focal loss Fold-1 max_auc=0.914 Fold-2 max_auc=0.914 Fold-3 max_auc=0.912 |
Fold=3 epochs=20 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
| 4 | EfficientNetB6 (imagenet) |
192x192 | 0.905 | 0.9302 | ✔️ | cosine_schedule Focal loss Fold-1 max_auc=0.903 Fold-2 max_auc=0.915 Fold-3 max_auc=0.899 |
Fold=3 epochs=20 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
| 5 | EfficientNetB6 (noisy-student) |
128x128 | 0.893 | - | ✔️ | cosine_schedule Focal loss Fold-1 max_auc=0.898 Fold-2 max_auc=0.907 Fold-3 max_auc=0.880 |
Fold=3 epochs=20 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
| 6 | EfficientNetB5 (noisy-student) |
512x512 | 0.931 | 0.9418 | ✔️ | cosine_schedule Focal loss Fold-1 max_auc=0.928 Fold-2 max_auc=0.931 Fold-3 max_auc=0.936 |
Fold=3 epochs=20 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
| 7 | EfficientNetB5 (noisy-student) |
384x384 | 0.915 | 0.9467 | ✔️ | Upsampling M1,M3,M4 cosine_schedule Focal loss Fold-1 max_auc=0.915 Fold-2 max_auc=0.925 Fold-3 max_auc=0.911 |
Fold=3 epochs=20 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
| 8 | EfficientNetB5 (noisy-student) |
256x256 | 0.918 | - | ✔️ | cosine_schedule Focal loss Fold-1 max_auc=0.914 Fold-2 max_auc=0.922 Fold-3 max_auc=0.918 |
Fold=3 epochs=20 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
| 9 | EfficientNetB5 (imagenet) |
192x192 | 0.906 | 0.9297 | ✔️ | cosine_schedule Focal loss Fold-1 max_auc=0.907 Fold-2 max_auc=0.912 Fold-3 max_auc=0.901 |
Fold=3 epochs=20 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
| 10 | EfficientNetB5 (noisy-student) |
128x128 | 0.892 | 0.9194 | ✔️ | cosine_schedule Focal loss Fold-1 max_auc=0.888 Fold-2 max_auc=0.903 Fold-3 max_auc=0.886 |
Fold=3 epochs=20 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
| 11 | EfficientNetB4 (noisy-student) |
512x512 | 0.920 | 0.9376 | ✔️ | cosine_schedule Focal loss Fold-1 max_auc=0.918 Fold-2 max_auc=0.927 Fold-3 max_auc=0.919 |
Fold=3 epochs=20 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
| 12 | EfficientNetB4 (noisy-student) |
384x384 | 0.913 | 0.9423 | ✔️ | Upsampling M1,M3,M4 cosine_schedule Focal loss Fold-1 max_auc=0.917 Fold-2 max_auc=0.917 Fold-3 max_auc=0.907 |
Fold=3 epochs=20 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
| 13 | EfficientNetB4 (noisy-student) |
256x256 | 0.913 | 0.9347 | ✔️ | cosine_schedule Focal loss Fold-1 max_auc=0.908 Fold-2 max_auc=0.922 Fold-3 max_auc=0.917 |
Fold=3 epochs=20 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
| 14 | EfficientNetB4 (noisy-student) |
192x192 | 0.905 | 0.9264 | ✔️ | cosine_schedule Focal loss Fold-1 max_auc=0.905 Fold-2 max_auc=0.911 Fold-3 max_auc=0.900 |
Fold=3 epochs=20 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
| 15 | EfficientNetB4 (noisy-student) |
128x128 | 0.895 | 0.9187 | ✔️ | cosine_schedule Focal loss Fold-1 max_auc=0.892 Fold-2 max_auc=0.906 Fold-3 max_auc=0.886 |
Fold=3 epochs=20 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
384 Series [S]
| No | model | Image-Size | CV-Score | LB-Score | External Data | Details | parameters |
|---|---|---|---|---|---|---|---|
| 1 | EfficientNetB6 (noisy-student) |
384x384 | 0.925 | 0.9399 | ✔️ | cosine_schedule Focal loss Fold-1 max_auc=0.916 Fold-2 max_auc=0.935 Fold-3 max_auc=0.925 |
Fold=3 epochs=20 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
| 2 | EfficientNetB5 (noisy-student) |
384x384 | 0.930 | 0.9434 | ✔️ | cosine_schedule Focal loss Fold-1 max_auc=0.92 Fold-2 max_auc=0.90 Fold-3 |
Fold=3 epochs=20 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
| 3 | EfficientNetB4 (noisy-student) |
384x384 | 0.921 | 0.9314 | ✔️ | cosine_schedule Focal loss Fold-1 max_auc=0.92 Fold-2 max_auc=0.924 Fold-3 max_auc=0.920 |
Fold=3 epochs=20 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
| 4 | EfficientNetB3 (noisy-student) |
384x384 | 0.915 | 0.9312 | ✔️ | cosine_schedule Focal loss Fold-1 max_auc=0.917 Fold-2 max_auc=0.921 Fold-3 max_auc=0.910 |
Fold=3 epochs=20 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
| 5 | EfficientNetB2 (noisy-student) |
384x384 | 0.915 | 0.9350 | ✔️ | cosine_schedule Focal loss Fold-1 max_auc=0.914 Fold-2 max_auc=0.924 Fold-3 max_auc=0.921 |
Fold=3 epochs=20 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
| 6 | EfficientNetB1 (noisy-student) |
384x384 | 0.919 | - | ✔️ | cosine_schedule Focal loss Fold-1 max_auc=0.914 Fold-2 max_auc=0.924 Fold-3 max_auc=0.921 |
Fold=3 epochs=20 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
| 7 | EfficientNetB0 (noisy-student) |
384x384 | 0.911 | - | ✔️ | cosine_schedule Focal loss Fold-1 max_auc=0.915 Fold-2 max_auc=0.918 Fold-3 max_auc=0.900 |
Fold=3 epochs=20 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
768 Series [P]
| No | model | Image-Size | CV-Score | LB-Score | External Data | Details | parameters |
|---|---|---|---|---|---|---|---|
| 1 | EfficientNetB6 (imagenet) |
768x768 | 0.927 | 0.9496 | ✔️ | cosine_schedule Focal loss Fold-1 max_auc=0.926 Fold-2 max_auc=0.926 Fold-3 max_auc=0.930 |
Fold=3 epochs=15 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
| 1_1 | EfficientNetB6 (imagenet) |
768x768 | 0.918 | 0.9368 | ✔️ | cosine_schedule BinaryCrossentropy Fold-1 max_auc=0.914 Fold-2 max_auc=0.929 Fold-3 max_auc=0.913 |
Fold=3 epochs=15 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
| 2 | EfficientNetB5 (imagenet) |
768x768 | 0.934 | 0.9462 | ✔️ | cosine_schedule Focal loss Fold-1 max_auc=0.936 Fold-2 max_auc=0.929 Fold-3 max_auc=0.937 |
Fold=3 epochs=15 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
| 3 | EfficientNetB4 (imagenet) |
768x768 | 0.929 | 0.9422 | ✔️ | cosine_schedule Focal loss Fold-1 max_auc=0.932 Fold-2 max_auc=0.929 Fold-3 max_auc=0.934 |
Fold=3 epochs=15 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
| 4 | EfficientNetB3 (imagenet) |
768x768 | 0.926 | 0.9454 | ✔️ | cosine_schedule Focal loss Fold-1 max_auc=0.929 Fold-2 max_auc=0.935 Fold-3 max_auc=0.923 |
Fold=3 epochs=20 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
| 5 | EfficientNetB2 (imagenet) |
768x768 | 0.924 | 0.9394 | ✔️ | cosine_schedule Focal loss Fold-1 max_auc=0.930 Fold-2 max_auc=0.919 Fold-3 max_auc=0.929 |
Fold=3 epochs=20 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
| 6 | EfficientNetB1 (imagenet) |
768x768 | 0.927 | 0.9371 | ✔️ | cosine_schedule Focal loss Fold-1 max_auc=0.929 Fold-2 max_auc=0.933 Fold-3 max_auc=0.922 |
Fold=3 epochs=20 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
| 7 | EfficientNetB0 (imagenet) |
768x768 | 0.923 | - | ✔️ | cosine_schedule Focal loss Fold-1 max_auc=0.920 Fold-2 max_auc=0.929 Fold-3 max_auc=0.924 |
Fold=3 epochs=20 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
B7 Series [E]
| No | model | Image-Size | CV-Score | LB-Score | External Data | Details | parameters |
|---|---|---|---|---|---|---|---|
| 1 | EfficientNetB7 (imagenet) |
768x768 | 0.937 | 0.9417 | ✔️ | cosine_schedule Focal loss Fold-1 max_auc=0.933 Fold-2 max_auc=0.947 |
Fold=2 epochs=20 TTA=20 INC2019 = [0,0] INC2018 = [1,1] |
| 2 | EfficientNetB7 (imagenet) |
512x512 | 0.934 | 0.9453 | ✔️ | cosine_schedule Focal loss Fold-1 max_auc=0.932 Fold-2 max_auc=0.934 Fold-3 max_auc=0.938 |
Fold=3 epochs=15 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
| 3 | EfficientNetB7 (imagenet) |
384x384 | 0.926 | 0.9424 | ✔️ | cosine_schedule Focal loss Fold-1 max_auc=0.921 Fold-2 max_auc=0.930 Fold-3 max_auc=0.929 |
Fold=3 epochs=15 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
| 4 | EfficientNetB7 (imagenet) |
256x256 | 0.921 | 0.9370 | ✔️ | cosine_schedule Focal loss Fold-1 max_auc=0.920 Fold-2 max_auc=0.925 Fold-3 max_auc=0.919 |
Fold=3 epochs=15 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
| 5 | EfficientNetB7 (imagenet) |
192x192 | 0.915 | 0.9310 | ✔️ | cosine_schedule Focal loss Fold-1 max_auc=0.912 Fold-2 max_auc=0.920 Fold-3 max_auc=0.913 |
Fold=3 epochs=15 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
| 6 | EfficientNetB7 (imagenet) |
128x128 | 0.893 | 0.9191 | ✔️ | cosine_schedule Focal loss Fold-1 max_auc=0.895 Fold-2 max_auc=0.907 Fold-3 max_auc=0.884 |
Fold=3 epochs=15 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
1024 Series [H]
| No | model | Image-Size | CV-Score | LB-Score | External Data | Details | parameters |
|---|---|---|---|---|---|---|---|
| 1 | EfficientNetB7 (imagenet) |
1024x1024 | - | - | ✔️ | cosine_schedule Focal loss Fold-1 max_auc=0.926 Fold-2 max_auc=0.926 Fold-3 max_auc=0.930 |
Fold=3 epochs=15 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
| 2 | EfficientNetB6 (imagenet) |
1024x1024 | 0.931 | 0.9423 | ✔️ | cosine_schedule Focal loss Fold-1 max_auc=0.938 Fold-2 max_auc=0.931 Fold-3 max_auc=0.929 |
Fold=3 epochs=15 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
| 3 | EfficientNetB5 (imagenet) |
1024x1024 | 0.932 | 0.9482 | ✔️ | cosine_schedule Focal loss Fold-1 max_auc=0.938 Fold-2 max_auc=0.931 Fold-3 max_auc=0.933 |
Fold=3 epochs=15 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
| 4 | EfficientNetB4 (imagenet) |
1024x1024 | 0.927 | 0.9405 | ✔️ | cosine_schedule Focal loss Fold-1 max_auc=0.935 Fold-2 max_auc=0.925 Fold-3 max_auc=0.921 |
Fold=3 epochs=15 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
| 5 | EfficientNetB3 (imagenet) |
1024x1024 | - | - | ✔️ | cosine_schedule Focal loss Fold-1 max_auc=0.926 Fold-2 max_auc=0.926 Fold-3 max_auc=0.930 |
Fold=3 epochs=15 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
| 6 | EfficientNetB2 (imagenet) |
1024x1024 | - | - | ✔️ | cosine_schedule Focal loss Fold-1 max_auc=0.926 Fold-2 max_auc=0.926 Fold-3 max_auc=0.930 |
Fold=3 epochs=15 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
| 6 | EfficientNetB1 (imagenet) |
1024x1024 | - | - | ✔️ | cosine_schedule Focal loss Fold-1 max_auc=0.926 Fold-2 max_auc=0.926 Fold-3 max_auc=0.930 |
Fold=3 epochs=15 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
| 6 | EfficientNetB0 (imagenet) |
1024x1024 | - | - | ✔️ | cosine_schedule Focal loss Fold-1 max_auc=0.926 Fold-2 max_auc=0.926 Fold-3 max_auc=0.930 |
Fold=3 epochs=15 TTA=20 INC2019 = [0,0,0] INC2018 = [1,1,1] |
New Seed [N]
| No | model | Image-Size | CV-Score | LB-Score | External Data | Details | parameters |
|---|---|---|---|---|---|---|---|
| 1 | EfficientNetB7 (imagenet) |
512x512 | 0.9377 | - | ✔️ | BCE Focal loss |
Fold=3 epochs=20 M3 CoutOut TTA=25 INC2019 = [0,0,0] INC2018 = [1,1,1] |
| 2 | EfficientNetB6 (imagenet) |
512x512 | 0.9345 | - | ✔️ | BCE Focal loss |
Fold=3 epochs=20 M3 CoutOut TTA=25 INC2019 = [0,0,0] INC2018 = [1,1,1] |
| 3 | EfficientNetB5 (imagenet) |
512x512 | 0.9393 | - | ✔️ | BCE Focal loss |
Fold=3 epochs=20 M3 CoutOut TTA=25 INC2019 = [0,0,0] INC2018 = [1,1,1] |
| 4 | EfficientNetB5 (imagenet) |
512x512 | 0.9334 | - | ✔️ | BCE Focal loss |
Fold=3 epochs=20 M3 CoutOut TTA=25 INC2019 = [0,0,0] INC2018 = [1,1,1] |
| 5 | EfficientNetB5 (imagenet) |
768x768 | 0.9383 | - | ✔️ | BCE Focal loss |
Fold=3 epochs=20 M3 CoutOut TTA=25 INC2019 = [0,0,0] INC2018 = [1,1,1] |