crop update
Browse files
README.md
CHANGED
|
@@ -158,7 +158,11 @@ python evaluate_external.py \
|
|
| 158 |
- **Single-source dataset** Performance on data from
|
| 159 |
other scanners, institutions, or populations is unknown and likely lower.
|
| 160 |
- **Cropped-ROI inputs.** The model expects nodule-cropped images like TN5000;
|
| 161 |
-
whole-frame ultrasound will be out of distribution.
|
|
|
|
|
|
|
|
|
|
|
|
|
| 162 |
- **Sensitivity gap at deployment threshold.** The threshold targets ≥0.95
|
| 163 |
sensitivity on validation; on the test set sensitivity was 0.904 — i.e. the
|
| 164 |
operating point does not perfectly transfer, and ~10% of malignant nodules
|
|
|
|
| 158 |
- **Single-source dataset** Performance on data from
|
| 159 |
other scanners, institutions, or populations is unknown and likely lower.
|
| 160 |
- **Cropped-ROI inputs.** The model expects nodule-cropped images like TN5000;
|
| 161 |
+
whole-frame ultrasound will be out of distribution. The model input is a region of interest
|
| 162 |
+
centered on the nodule that includes the nodule together with a margin of adjacent tissue.
|
| 163 |
+
It is not a tight crop that excludes all surrounding parenchyma. At deployment,
|
| 164 |
+
an operator selects the region of interest, so region selection is a human in the loop step,
|
| 165 |
+
and inter operator variability in region selection may affect model output.
|
| 166 |
- **Sensitivity gap at deployment threshold.** The threshold targets ≥0.95
|
| 167 |
sensitivity on validation; on the test set sensitivity was 0.904 — i.e. the
|
| 168 |
operating point does not perfectly transfer, and ~10% of malignant nodules
|