def compute_metrics(truth_file_stream, prediction_file_stream) -> Dict[str, Dict[str, float]]:
truth_probabilities = parse_csv(truth_file_stream, CATEGORIES)
prediction_probabilities = parse_csv(prediction_file_stream, CATEGORIES)
exclude_rows(truth_probabilities, EXCLUDE_LABELS)
exclude_rows(prediction_probabilities, EXCLUDE_LABELS)
validate_rows(truth_probabilities, prediction_probabilities)
sort_rows(truth_probabilities)
sort_rows(prediction_probabilities)
scores: Dict[str, Dict[str, float]] = {}
for category in CATEGORIES:
truth_category_probabilities: pd.Series = truth_probabilities[category]
prediction_category_probabilities: pd.Series = prediction_probabilities[category]
truth_binary_values: pd.Series = truth_category_probabilities.gt(0.5)
prediction_binary_values: pd.Series = prediction_category_probabilities.gt(0.5)
category_cm = create_binary_confusion_matrix(
truth_binary_values=truth_binary_values.to_numpy(),
prediction_binary_values=prediction_binary_values.to_numpy(),
name=category,
)
scores[category] = {
'accuracy': metrics.binary_accuracy(category_cm),
'sensitivity': metrics.binary_sensitivity(category_cm),
'specificity': metrics.binary_specificity(category_cm),
'dice': metrics.binary_dice(category_cm),
'ppv': metrics.binary_ppv(category_cm),
'npv': metrics.binary_npv(category_cm),
'auc': metrics.auc(truth_category_probabilities, prediction_category_probabilities),
'auc_sens_80': metrics.auc_above_sensitivity(
truth_category_probabilities, prediction_category_probabilities, 0.80
),
'ap': metrics.average_precision(
truth_category_probabilities, prediction_category_probabilities
),
}
# Compute averages for all per-category metrics
per_category_metrics: ValuesView[str] = next(iter(scores.values())).keys()
scores['macro_average'] = {
metric: float(np.mean([scores[category][metric] for category in CATEGORIES]))
for metric in per_category_metrics
}
# Compute multi-category aggregate metrics
scores['aggregate'] = {
'balanced_accuracy': metrics.balanced_multiclass_accuracy(
truth_probabilities, prediction_probabilities
)
}
return scores
def from_stream(
cls, truth_file_stream: TextIO, prediction_file_stream: TextIO
):
truth_probabilities, truth_weights = parse_truth_csv(truth_file_stream)
categories = truth_probabilities.columns
prediction_probabilities = parse_csv(prediction_file_stream, categories)
validate_rows(truth_probabilities, prediction_probabilities)
sort_rows(truth_probabilities)
sort_rows(prediction_probabilities)
score = cls(truth_probabilities, prediction_probabilities, truth_weights)
return score
def test_validate_rows_missing_images(categories):
truth_probabilities = pd.DataFrame(
[[1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0],
[0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0]],
index=['ISIC_0000123', 'ISIC_0000124'],
columns=categories,
)
prediction_probabilities = pd.DataFrame(
[[1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0],
[0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0]],
index=['ISIC_0000123', 'ISIC_0000124'],
columns=categories,
)
load_csv.validate_rows(truth_probabilities, prediction_probabilities)
truth_probabilities = pd.DataFrame(
[[1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0],
[0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0]],
index=['ISIC_0000123', 'ISIC_0000124'],
columns=categories,
)
prediction_probabilities = pd.DataFrame(
[[1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0]],
index=['ISIC_0000123'],
columns=categories)
with pytest.raises(ScoreException) as exc_info:
load_csv.validate_rows(truth_probabilities, prediction_probabilities)
assert "Missing images in CSV: ['ISIC_0000124']." == str(exc_info.value)
truth_probabilities = pd.DataFrame(
[[1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0],
[0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0]],
index=['ISIC_0000123', 'ISIC_0000124'],
columns=categories,
)
prediction_probabilities = pd.DataFrame(
[[1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0]],
index=['ISIC_0000120'],
columns=categories)
with pytest.raises(ScoreException) as exc_info:
load_csv.validate_rows(truth_probabilities, prediction_probabilities)
assert "Missing images in CSV: ['ISIC_0000123', 'ISIC_0000124']." == str(
exc_info.value)
truth_probabilities = pd.DataFrame(
[[1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0],
[0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0]],
index=['ISIC_0000123', 'ISIC_0000124'],
columns=categories,
)
prediction_probabilities = pd.DataFrame(
[[1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0],
[0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0]],
index=['ISIC_0000123', 'ISIC_0000125'],
columns=categories,
)
with pytest.raises(ScoreException) as exc_info:
load_csv.validate_rows(truth_probabilities, prediction_probabilities)
assert "Missing images in CSV: ['ISIC_0000124']." == str(exc_info.value)
def test_validate_rows_extra_images(categories):
truth_probabilities = pd.DataFrame(
[[1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0]], index=['ISIC_0000123'], columns=categories
)
prediction_probabilities = pd.DataFrame(
[
[1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0],
[0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0],
[0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0],
],
index=['ISIC_0000123', 'ISIC_0000126', 'ISIC_0000127'],
columns=categories,
)
with pytest.raises(ScoreException) as exc_info:
load_csv.validate_rows(truth_probabilities, prediction_probabilities)
assert 'Extra images in CSV: [\'ISIC_0000126\', \'ISIC_0000127\'].' == str(exc_info.value)
def compute_metrics(truth_file_stream, prediction_file_stream) -> ScoresType:
truth_probabilities, truth_weights = parse_truth_csv(truth_file_stream)
categories = truth_probabilities.columns
prediction_probabilities = parse_csv(prediction_file_stream, categories)
validate_rows(truth_probabilities, prediction_probabilities)
sort_rows(truth_probabilities)
sort_rows(prediction_probabilities)
scores: ScoresType = {}
for category in categories:
truth_category_probabilities: pd.Series = truth_probabilities[category]
prediction_category_probabilities: pd.Series = prediction_probabilities[
category]
truth_binary_values: pd.Series = truth_category_probabilities.gt(0.5)
prediction_binary_values: pd.Series = prediction_category_probabilities.gt(
0.5)
category_cm = create_binary_confusion_matrix(
truth_binary_values=truth_binary_values.to_numpy(),
prediction_binary_values=prediction_binary_values.to_numpy(),
weights=truth_weights.score_weight.to_numpy(),
name=category,
)
scores[category] = {
'accuracy':
metrics.binary_accuracy(category_cm),
'sensitivity':
metrics.binary_sensitivity(category_cm),
'specificity':
metrics.binary_specificity(category_cm),
'dice':
metrics.binary_dice(category_cm),
'ppv':
metrics.binary_ppv(category_cm),
'npv':
metrics.binary_npv(category_cm),
'auc':
metrics.auc(
truth_category_probabilities,
prediction_category_probabilities,
truth_weights.score_weight,
),
'auc_sens_80':
metrics.auc_above_sensitivity(
truth_category_probabilities,
prediction_category_probabilities,
truth_weights.score_weight,
0.80,
),
'ap':
metrics.average_precision(
truth_category_probabilities,
prediction_category_probabilities,
truth_weights.score_weight,
),
'roc':
metrics.roc(
truth_category_probabilities,
prediction_category_probabilities,
truth_weights.score_weight,
),
}
# Compute averages for all per-category metrics
per_category_metrics: KeysView[str] = next(iter(scores.values())).keys()
scores['macro_average'] = {
metric:
float(np.mean([scores[category][metric] for category in categories]))
for metric in per_category_metrics if metric != 'roc'
}
# Compute multi-category aggregate metrics
scores['aggregate'] = {
'balanced_accuracy':
metrics.balanced_multiclass_accuracy(truth_probabilities,
prediction_probabilities,
truth_weights.score_weight)
}
scores['overall'] = scores['aggregate']['balanced_accuracy']
scores['validation'] = metrics.balanced_multiclass_accuracy(
truth_probabilities, prediction_probabilities,
truth_weights.validation_weight)
return scores