Step 1: Senescence Patch Extraction

The code extracts overlapping nucleus and non-nucleus patches from cellular senescence images.

The extracted patches are later used for:

• Contrastive learning with encoder training
• Ranking patches
• Learning nucleus-cytoplasm relationships

Cell Lines

• HFF
• HMC3

Senescence Conditions

• H2O2
• Dox
• Pal

Users can change the folder names and paths according to their own dataset structure.

---

Input

The patch extraction code reads:

• Control images
• Senescence images

and generates:

• Overlapping nucleus patches
• Overlapping non-nucleus patches

---

Input Paths

Update the folder paths before running the code:

control_path = "/path/to/control_images"
senescence_path = "/path/to/senescence_images"
output_path = "/path/to/output_folder"

Example:

control_path = "/data/HFF/control"
senescence_path = "/data/HFF/h2o2"
output_path = "/data/HFF/output"

---

Run

Open and run:

Patch_extractor.ipynb

Step 2: Contrastive Learning with E-SupConLoss

After patch extraction, the extracted nucleus and non-nucleus patches are used for contrastive learning.

This step trains an encoder with an MLP projection head using E-SupConLoss. The goal is to learn meaningful patch-level embeddings for control and senescence samples.

The training uses:

• Control nucleus patches
• Control non-nucleus patches
• Senescence nucleus patches
• Senescence non-nucleus patches

The trained encoder is later used for patch ranking and interaction learning.

Input

This step uses the extracted patches from Step 1.

---

Run

python training_Esupconloss.py

---

Output

esupconloss_encoder.pth

Step 3: Ranking Patches

After contrastive learning, the saved encoder from Step 2 is used to rank nucleus and non-nucleus patches.

This step trains a patch-level classification model using the frozen encoder learned from E-SupConLoss training.

The model learns to classify whether a patch belongs to:

• Control
• Senescence

The trained model is then used to rank the most important nucleus and non-nucleus patches.

---

Input

This step uses:

• Extracted patches from Step 1
• Saved encoder from Step 2

---

Run

python Ranking_patches.py

---

Output

encoder_mlp_patch_classifier.pth

Step 4: Senescence Scoring with Cross-Attention

After ranking the patches, the top-k nucleus and non-nucleus patches are selected for senescence prediction.

This step uses a cross-attention network to learn relationships between nucleus and non-nucleus patches and generates a senescence probability score for each image.

The workflow includes:

• Patch extraction
• Patch ranking
• Top-k patch selection
• Cross-attention interaction learning
• Senescence probability prediction

---

Input

This step uses:

• Control and senescence images

---

Run

python cross_attention_senescence_scoring.py

---

Output

The model generates a probability score indicating whether the image belongs to:

• Control
• Senescence