LookBack / README.md
fvewa's picture
Upload folder using huggingface_hub (part 2)
3e80f2d verified
|
Raw History Blame Contribute Delete
5.38 kB
# NBME Lung AI Anchoring — Dataset
Gaze recordings, radiologist annotations, AI predictions, and imaging data from a within-subject three-condition study of AI-induced attentional anchoring in 3D CT lung nodule reading (16 radiologists × 60 LIDC-IDRI cases × 3 conditions).
## Repository contents
```
Collections/ # Per-session gaze and annotation data
├── R{1-16}_Set{1,2}_R1/ # First run: Exp2 (Concurrent CAD)
│ ├── Trial1.csv # Gaze time series (6 columns, no Phase)
│ └── Trial1.json # Reader bounding-box annotations for Exp2
│
└── R{1-16}_Set{1,2}_R2/ # Second run: Exp3 (Gaze-guided second look)
├── Trial1.csv # Gaze time series (7 columns, with Phase)
├── Trial1_firstread.json # Annotations after independent first read (also used as Exp1)
└── Trial1_secondlook.json # Annotations added after gaze-guided second-look prompt
images_bbox/
└── sets/
├── set1/ # 30 cases
│ ├── ai_nodules.json # AI (RetinaNet) predicted bounding boxes per case
│ ├── gt_nodules.json # Ground-truth nodule bounding boxes per case
│ ├── vol/ # CT volumes (.nii.gz), one per case
│ ├── gt/ # GT nodule masks
│ ├── lung_mask/ # Lung field masks
│ └── pred/ # Per-case AI prediction files
└── set2/ # 30 cases, same structure
```
## Reader anonymization
Original participant names have been replaced with anonymous IDs R1–R16. R12 participated in Set 1 only; all other readers completed both sets.
## Session naming and mapping to experimental conditions
`R{reader}_Set{1,2}_R{1,2}` — e.g. `R3_Set1_R2` is reader R3, Set 1, Run 2.
- `Set1` = 30 CT cases; `Set2` = 30 CT cases
- `R1` (first run) is always **Exp2 (Concurrent CAD)**: the reader read each case with AI-suggested bounding boxes shown throughout.
- `R2` (second run) is **Exp3 (Gaze-guided second look)**, which is a two-phase session recorded in a single file set. The first-read phase of `R2` is additionally used as **Exp1 (Independent reading)** in all analyses that require a no-AI baseline from the same reader on the same cases.
Phase extraction:
| Analysis target | Gaze (Trial1.csv) | Annotations |
|---|---|---|
| **Exp1 (Independent)** | `R2` session, rows where `Phase == "first_read"` | `Trial1_firstread.json` from `R2` session |
| **Exp2 (Concurrent CAD)** | `R1` session (entire CSV; no Phase column) | `Trial1.json` from `R1` session |
| **Exp3 (Gaze-guided second look)** | `R2` session, both `Phase == "first_read"` and `Phase == "second_look"` rows combined | `Trial1_firstread.json` + `Trial1_secondlook.json` from `R2` session (combined; de-duplicate overlapping boxes) |
Notes:
- `R1` session CSVs have 6 columns (no `Phase`) because Exp2 is a single uninterrupted read.
- `R2` session CSVs have 7 columns, the last being `Phase ∈ {"first_read", "second_look"}`.
- Every `R2` session contains both `Trial1_firstread.json` and `Trial1_secondlook.json`. The `secondlook` file lists annotations added after the gaze-guided second-look prompt; some may overlap `firstread` boxes and should be de-duplicated when forming Exp3 annotations.
- R12 did not have an Exp2 run for Set 2 (no `R12_Set2_R1` session). `R12_Set2_R2` is present but was excluded from all analyses in the companion paper due to data-quality issues (see the paper's analysis-populations table).
## File formats
### Trial1.csv (gaze recording)
Columns:
- `Time`: HH:MM:SS wall-clock timestamp (local to recording session)
- `Gaze_Location_X/Y`: screen coordinates in pixels (display 2560×1080)
- `Slice Number`: CT axial slice index (0-based) at the moment of fixation
- `Duration`: fixation duration (sampling-rate dependent; typically ~1 at 50–60 Hz)
- `File Name`: relative path to the viewed CT volume (`set{N}/vol/LIDC-IDRI-XXXX_vol.nii.gz`)
- `Phase` (R2 sessions only): `first_read` or `second_look`
### Trial1.json / Trial1_firstread.json / Trial1_secondlook.json (annotations)
Dict mapping `LIDC-IDRI-XXXX` → list of bounding boxes.
Each bounding box: `[z1, y1, x1, z2, y2, x2]` where `z` indexes axial slices and `y, x` are in pixel coordinates relative to the CT volume.
### ai_nodules.json / gt_nodules.json
Dict mapping `LIDC-IDRI-XXXX` → list of bounding boxes in the same format.
### Volumes
NIfTI `.nii.gz` CT volumes derived from LIDC-IDRI. Original DICOM data is publicly available via The Cancer Imaging Archive (TCIA): https://www.cancerimagingarchive.net/collection/lidc-idri/
## Citation and license
Please cite the associated paper when using this dataset. The 60 CT cases are derived from LIDC-IDRI, which is distributed under its own license; please also cite LIDC-IDRI (Armato et al., *Medical Physics* 2011).
Dataset license: CC-BY 4.0 (see `LICENSE`), applied to the gaze recordings, reader annotations, AI predictions, and derived masks. LIDC-IDRI source imagery retains its original terms.