![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F11837581%2Fd1aff4084ad3b318e2a6350a4869fca3%2FScreenshot%202023-10-16%20at%2009.55.53.png?generation=1697440072235399&alt=media\")
I made a rather interesting approach to classify extravasation and want to share it with you.
\nMy idea would have been impossible to implement without the bounding box dataset from Ian Pan . Thanks a lot for his work.
\nUsing this dataset, I trained YOLO detection and collected my โtimeโ series dataset. The idea was to collect confidence and area of โโthe bounding boxes. Walking through the sorted scans of the axial plane, I collected model predictions into my dataset.
\n
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F11837581%2F6235a05af397f1b3e2e0e42c01c9585e%2FScreenshot%202023-10-16%20at%2010.08.53.png?generation=1697440146957921&alt=media\")
When assembling the dataset, I also experimented with TTA, but as practice has shown, this did not bring a big increase in quality, but took 4 times more time for inference.
\nAfter that, I started generating features for these series. After many attempts, I came to the conclusion that the simplest features, such as std, mean, median, etc. There were already enough of them for the optimal metric. I couldnโt separate the classes more clearly.
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F11837581%2Fc33d7d159235bb1375e68bf64315da5d%2FScreenshot%202023-10-16%20at%2010.11.35.png?generation=1697440490402173&alt=media\")
Using these features, I experimented with various models, but settled on the logistic regression ensemble and lmblearn Balanced Random Forest Classifier.
\nI made stratified cross-validation.
\nScore = Metric(label = 5)\nval_scores = []\n\n i range(N_SPLITS):\n X_train = df[feature_cols][df.fold != i]\n y_train = df[df.fold != i].label\n X_val = df[feature_cols][df.fold == i]\n y_val = df[df.fold == i].label\n\n LR = LogisticRegression(=21,\n =,\n class_weight={0 : 1, 1 : 6},\n =,\n =0.9)\n\n BRF = BalancedRandomForestClassifier(=100,\n =,\n =None,\n =2,\n =1,\n =0.,\n =,\n =None,\n =0.,\n =,\n =,\n =,\n =,\n =21,\n =0,\n =,\n class_weight={0 : 1, 1 : 5},\n =0.,\n =None\n )\n\n fit_LR = LR.fit(X_train, y_train)\n fit_BRF = BRF.fit(X_train, y_train)\n\n pred = np.array(0.5 * fit_BRF.predict(X_val) + 0.5 * fit_LR.predict(X_val), dtype = np.uint8)\n f1 = f1_score(y_val, pred)\n\n = np.array(y_val)\n pred_LR = np.array(fit_LR.predict_proba(X_val))\n pred_BRF = np.array(fit_BRF.predict_proba(X_val))\n pred = 0.5 * pred_BRF + 0.5 * pred_LR\n val_score = Score.get_score(, pred)\n val_scores.append(val_score)\n\n (f)\n (f)\n (f)\n ()\n\n(f)\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F11837581%2F4c45ae6d9091e8491ed5ad6553e1a396%2FScreenshot%202023-10-16%20at%2010.44.42.png?generation=1697442310307315&alt=media\")
This approach showed 0.02 better logloss than the best statistical approach.
\nIt seems that the idea can be improved, for example by collecting better data or generating more suitable features.
\nThank you for your attention, I look forward to your criticism and suggestions.
\nI'm waiting for your questions.
<3
", "messages": [ { "id": 2484072, "postDate": "2023-10-16T07:50:15.950Z", "content": "I made a rather interesting approach to classify extravasation and want to share it with you.
\nMy idea would have been impossible to implement without the bounding box dataset from Ian Pan . Thanks a lot for his work.
\nUsing this dataset, I trained YOLO detection and collected my โtimeโ series dataset. The idea was to collect confidence and area of โโthe bounding boxes. Walking through the sorted scans of the axial plane, I collected model predictions into my dataset.
\n
When assembling the dataset, I also experimented with TTA, but as practice has shown, this did not bring a big increase in quality, but took 4 times more time for inference.
\nAfter that, I started generating features for these series. After many attempts, I came to the conclusion that the simplest features, such as std, mean, median, etc. There were already enough of them for the optimal metric. I couldnโt separate the classes more clearly.
\n
Using these features, I experimented with various models, but settled on the logistic regression ensemble and lmblearn Balanced Random Forest Classifier.
\nI made stratified cross-validation.
\nScore = Metric(label = 5)\nval_scores = []\n\n i range(N_SPLITS):\n X_train = df[feature_cols][df.fold != i]\n y_train = df[df.fold != i].label\n X_val = df[feature_cols][df.fold == i]\n y_val = df[df.fold == i].label\n\n LR = LogisticRegression(=21,\n =,\n class_weight={0 : 1, 1 : 6},\n =,\n =0.9)\n\n BRF = BalancedRandomForestClassifier(=100,\n =,\n =None,\n =2,\n =1,\n =0.,\n =,\n =None,\n =0.,\n =,\n =,\n =,\n =,\n =21,\n =0,\n =,\n class_weight={0 : 1, 1 : 5},\n =0.,\n =None\n )\n\n fit_LR = LR.fit(X_train, y_train)\n fit_BRF = BRF.fit(X_train, y_train)\n\n pred = np.array(0.5 * fit_BRF.predict(X_val) + 0.5 * fit_LR.predict(X_val), dtype = np.uint8)\n f1 = f1_score(y_val, pred)\n\n = np.array(y_val)\n pred_LR = np.array(fit_LR.predict_proba(X_val))\n pred_BRF = np.array(fit_BRF.predict_proba(X_val))\n pred = 0.5 * pred_BRF + 0.5 * pred_LR\n val_score = Score.get_score(, pred)\n val_scores.append(val_score)\n\n (f)\n (f)\n (f)\n ()\n\n(f)\n
This approach showed 0.02 better logloss than the best statistical approach.
\nIt seems that the idea can be improved, for example by collecting better data or generating more suitable features.
\nThank you for your attention, I look forward to your criticism and suggestions.
\nI'm waiting for your questions.
<3
", "rawMarkdown": "I made a rather interesting approach to classify extravasation and want to share it with you.\n\nMy idea would have been impossible to implement without the bounding box [dataset](https://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/discussion/441402) from [Ian Pan]()(https://www.kaggle.com/vaillant) . Thanks a lot for his work.\n\nUsing this dataset, I trained YOLO detection and collected my โtimeโ series dataset. The idea was to collect confidence and area of โโthe bounding boxes. Walking through the sorted scans of the axial plane, I collected model predictions into my dataset.\n\n# ๐๐ก๐ข๐ฌ ๐ข๐ฌ ๐ฐ๐ก๐๐ญ ๐ญ๐ก๐ โ๐ญ๐ข๐ฆ๐โ ๐ฌ๐๐ซ๐ข๐๐ฌ ๐ฅ๐จ๐จ๐ค๐๐ ๐ฅ๐ข๐ค๐:\n\n## ๐๐ฑ๐ญ๐ซ๐๐ฏ๐๐ฌ๐๐ญ๐ข๐จ๐ง\n\n\n\n## ๐๐จ ๐๐ฑ๐ญ๐ซ๐๐ฏ๐๐ฌ๐๐ญ๐ข๐จ๐ง\n\n\n# ๐๐๐ญ๐ ๐ฉ๐ซ๐๐ฉ๐๐ซ๐ข๐ง๐ :\nWhen assembling the dataset, I also experimented with TTA, but as practice has shown, this did not bring a big increase in quality, but took 4 times more time for inference.\n\n# ๐ ๐๐๐ญ๐ฎ๐ซ๐ ๐๐ง๐ ๐ข๐ง๐๐๐ซ๐ข๐ง๐ :\nAfter that, I started generating features for these series. After many attempts, I came to the conclusion that the simplest features, such as std, mean, median, etc. There were already enough of them for the optimal metric. I couldnโt separate the classes more clearly.\n\n\n# ๐๐ฅ๐๐ฌ๐ฌ๐ข๐๐ข๐๐๐ญ๐ข๐จ๐ง ๐ฆ๐จ๐๐๐ฅ๐ฌ:\nUsing these features, I experimented with various models, but settled on the logistic regression ensemble and [lmblearn Balanced Random Forest Classifier](https://imbalanced-learn.org/stable/references/generated/imblearn.ensemble.BalancedRandomForestClassifier.html).\n\nI made stratified cross-validation.\n```\nScore = Metric(label = 5)\nval_scores = []\n\nfor i in range(N_SPLITS):\n X_train = df[feature_cols][df.fold != i]\n y_train = df[df.fold != i].label\n X_val = df[feature_cols][df.fold == i]\n y_val = df[df.fold == i].label\n \n LR = LogisticRegression(random_state=21,\n penalty='elasticnet',\n class_weight={0 : 1, 1 : 6},\n solver=\"saga\",\n l1_ratio=0.9)\n\n BRF = BalancedRandomForestClassifier(n_estimators=100,\n criterion=\"gini\",\n max_depth=None,\n min_samples_split=2,\n min_samples_leaf=1,\n min_weight_fraction_leaf=0.,\n max_features='sqrt',\n max_leaf_nodes=None,\n min_impurity_decrease=0.,\n bootstrap=True,\n oob_score=False,\n sampling_strategy=\"auto\",\n replacement=False,\n random_state=21,\n verbose=0,\n warm_start=False,\n class_weight={0 : 1, 1 : 5},\n ccp_alpha=0.,\n max_samples=None\n )\n\n fit_LR = LR.fit(X_train, y_train)\n fit_BRF = BRF.fit(X_train, y_train)\n \n pred = np.array(0.5 * fit_BRF.predict(X_val) + 0.5 * fit_LR.predict(X_val), dtype = np.uint8)\n f1 = f1_score(y_val, pred)\n \n true = np.array(y_val)\n pred_LR = np.array(fit_LR.predict_proba(X_val))\n pred_BRF = np.array(fit_BRF.predict_proba(X_val))\n pred = 0.5 * pred_BRF + 0.5 * pred_LR\n val_score = Score.get_score(true, pred)\n val_scores.append(val_score)\n \n print(f'fold: {i}')\n print(f'f1 val score: {f1}')\n print(f'w_logloss val score: {val_score}')\n print()\n \nprint(f\"mean val score: {np.mean(val_scores)} +- {2 * np.std(val_scores)}\")\n```\n\n# ๐๐ง๐ ๐ ๐จ๐ญ ๐ญ๐ก๐ ๐๐จ๐ฅ๐ฅ๐จ๐ฐ๐ข๐ง๐ ๐ซ๐๐ฌ๐ฎ๐ฅ๐ญ๐ฌ:\n\n\n\n\nThis approach showed 0.02 better logloss than the best statistical approach.\nIt seems that the idea can be improved, for example by collecting better data or generating more suitable features.\nThank you for your attention, I look forward to your criticism and suggestions.\nI'm waiting for your questions.\n\n<3\n\n\n", "votes": 13 } ], "comments": [], "raw_markdown_by_id": { "2484072": "I made a rather interesting approach to classify extravasation and want to share it with you.\n\nMy idea would have been impossible to implement without the bounding box [dataset](https://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/discussion/441402) from [Ian Pan]()(https://www.kaggle.com/vaillant) . Thanks a lot for his work.\n\nUsing this dataset, I trained YOLO detection and collected my โtimeโ series dataset. The idea was to collect confidence and area of โโthe bounding boxes. Walking through the sorted scans of the axial plane, I collected model predictions into my dataset.\n\n# ๐๐ก๐ข๐ฌ ๐ข๐ฌ ๐ฐ๐ก๐๐ญ ๐ญ๐ก๐ โ๐ญ๐ข๐ฆ๐โ ๐ฌ๐๐ซ๐ข๐๐ฌ ๐ฅ๐จ๐จ๐ค๐๐ ๐ฅ๐ข๐ค๐:\n\n## ๐๐ฑ๐ญ๐ซ๐๐ฏ๐๐ฌ๐๐ญ๐ข๐จ๐ง\n\n\n\n## ๐๐จ ๐๐ฑ๐ญ๐ซ๐๐ฏ๐๐ฌ๐๐ญ๐ข๐จ๐ง\n\n\n# ๐๐๐ญ๐ ๐ฉ๐ซ๐๐ฉ๐๐ซ๐ข๐ง๐ :\nWhen assembling the dataset, I also experimented with TTA, but as practice has shown, this did not bring a big increase in quality, but took 4 times more time for inference.\n\n# ๐ ๐๐๐ญ๐ฎ๐ซ๐ ๐๐ง๐ ๐ข๐ง๐๐๐ซ๐ข๐ง๐ :\nAfter that, I started generating features for these series. After many attempts, I came to the conclusion that the simplest features, such as std, mean, median, etc. There were already enough of them for the optimal metric. I couldnโt separate the classes more clearly.\n\n\n# ๐๐ฅ๐๐ฌ๐ฌ๐ข๐๐ข๐๐๐ญ๐ข๐จ๐ง ๐ฆ๐จ๐๐๐ฅ๐ฌ:\nUsing these features, I experimented with various models, but settled on the logistic regression ensemble and [lmblearn Balanced Random Forest Classifier](https://imbalanced-learn.org/stable/references/generated/imblearn.ensemble.BalancedRandomForestClassifier.html).\n\nI made stratified cross-validation.\n```\nScore = Metric(label = 5)\nval_scores = []\n\nfor i in range(N_SPLITS):\n X_train = df[feature_cols][df.fold != i]\n y_train = df[df.fold != i].label\n X_val = df[feature_cols][df.fold == i]\n y_val = df[df.fold == i].label\n \n LR = LogisticRegression(random_state=21,\n penalty='elasticnet',\n class_weight={0 : 1, 1 : 6},\n solver=\"saga\",\n l1_ratio=0.9)\n\n BRF = BalancedRandomForestClassifier(n_estimators=100,\n criterion=\"gini\",\n max_depth=None,\n min_samples_split=2,\n min_samples_leaf=1,\n min_weight_fraction_leaf=0.,\n max_features='sqrt',\n max_leaf_nodes=None,\n min_impurity_decrease=0.,\n bootstrap=True,\n oob_score=False,\n sampling_strategy=\"auto\",\n replacement=False,\n random_state=21,\n verbose=0,\n warm_start=False,\n class_weight={0 : 1, 1 : 5},\n ccp_alpha=0.,\n max_samples=None\n )\n\n fit_LR = LR.fit(X_train, y_train)\n fit_BRF = BRF.fit(X_train, y_train)\n \n pred = np.array(0.5 * fit_BRF.predict(X_val) + 0.5 * fit_LR.predict(X_val), dtype = np.uint8)\n f1 = f1_score(y_val, pred)\n \n true = np.array(y_val)\n pred_LR = np.array(fit_LR.predict_proba(X_val))\n pred_BRF = np.array(fit_BRF.predict_proba(X_val))\n pred = 0.5 * pred_BRF + 0.5 * pred_LR\n val_score = Score.get_score(true, pred)\n val_scores.append(val_score)\n \n print(f'fold: {i}')\n print(f'f1 val score: {f1}')\n print(f'w_logloss val score: {val_score}')\n print()\n \nprint(f\"mean val score: {np.mean(val_scores)} +- {2 * np.std(val_scores)}\")\n```\n\n# ๐๐ง๐ ๐ ๐จ๐ญ ๐ญ๐ก๐ ๐๐จ๐ฅ๐ฅ๐จ๐ฐ๐ข๐ง๐ ๐ซ๐๐ฌ๐ฎ๐ฅ๐ญ๐ฌ:\n\n\n\n\nThis approach showed 0.02 better logloss than the best statistical approach.\nIt seems that the idea can be improved, for example by collecting better data or generating more suitable features.\nThank you for your attention, I look forward to your criticism and suggestions.\nI'm waiting for your questions.\n\n<3\n\n\n" } }