数据集获取

git clone https://github.com/juliadietlmeier/CF-OCT-segmentation.git

curl -L -o repo.zip https://github.com/juliadietlmeier/CF-OCT-segmentation/archive/refs/heads/main.zip
unzip repo.zip

源站 README 摘录（使用方式）

CF-OCT-segmentation

This repository contains the code for the semantic segmentation and quantification of cochlear fibrosis (CF) in a novel Cochlear Fibrosis OCT dataset (animal models). We hope that our findings will advance future studies on exploring the relationship between cochlear fibrosis and residual hearing loss, development of cochlear implants (CI) or treatment of EAS-patients.
We benchmark different Keras segmentation models on our new Cochlear Fibrosis OCT dataset:
Proposed 2D-OCT-UNET
Fig. 1. Block diagram of the proposed 2D-OCT-UNET architecture for the multiclass OCT segmentation used in this work. The very deep 2D-
OCT-UNET processes two-dimensional OCT slices and consists of seven encoder-decoder blocks with skip connections. The numbers above the
encoder, bottleneck and decoder blocks indicate the number of filters in the convolutional Conv2D layers. GN stands for Group Normalization layers
with the number of groups parameter ng = 2. The input resolution of the UNET is set to be 1024 × 1024 pixels. We include Dropout(0.1) layers
only in the decoder. The number of filters in the last Conv2D layer is equal to the nc = 4 (number of classes).
Other models included are: VGG16-UNET, UEfficientNet, SegFormer and MST-DeepLabv3
To benchmark the SAM (Segment Anything Model) on the Cochlear OCT dataset we refer to the Github repository:
https://github.com/mazurowski-lab/finetune-SAM

Cochlear Fibrosis OCT Dataset

The open-source dataset can be dowloaded from the Open Science Framework (OSF)
Associated OSF project
https://osf.io/cghn7
Registration DOI
https://doi.org/10.17605/OSF.IO/WB5FS
Fig. 2. Cochlear Fibrosis OCT dataset samples and the corresponding ground truth annotations from five annotated volumes. The CI/Track class is depicted
in red, the Fibrosis class in green and the ST/Free Space class in blue.

Fig. 3. Samples from qualitative results are provided for each volume. First two rows: OCTV1L, Third row: OCTV7L, Fourth row: OCTV9L, Fifth
row: OCTV10L and the Sixth row: OCTV11L. By comparing to the Ground Truth (GT) it can be inferred that the 2D-OCT-UNET model performed the
best and the SAM (prompt-free auto-mode implementation with adapters) model qualitatively performed the worst. As pre

数据加载示例（图像类）

from PIL import Image
import glob, os

files = (glob.glob(os.path.join(path, "**", "*.png"), recursive=True)
       + glob.glob(os.path.join(path, "**", "*.jpg"), recursive=True)
       + glob.glob(os.path.join(path, "**", "*.tif"), recursive=True))
print("图像文件数:", len(files))
img = Image.open(files[0]); print("尺寸/模式:", img.size, img.mode)

# torchvision Dataset 方式:
# from torchvision import datasets
# ds = datasets.ImageFolder(path)  # 要求 子目录=类别

目录组织与标注格式以源站说明和下载后实际文件为准。

完整仓库：github.com/juliadietlmeier/CF-OCT-segmentation