setting,contrastive_weight,regression_weight,checkpoint,mae,rmse,pearson,spearman,pet_to_suvr_r1,pet_to_suvr_r5,suvr_to_pet_r1,suvr_to_pet_r5,mrr_pet_to_suvr,mrr_suvr_to_pet,median_rank_pet_to_suvr,median_rank_suvr_to_pet,top5_high,top5_low,interpretation
Regression only,0.0,1.0,runs/foundation/remap_pet_layer4_regonly_best.pt,0.0549,0.0710,0.9538,0.9535,0.0065,0.0392,0.0065,0.0131,0.0400,0.0315,76,83,0.6732,0.8000,best numeric SUVR prediction but no PET-SUVR alignment
Weak contrastive,0.1,1.0,runs/foundation/remap_pet_layer4_cw01_best.pt,0.0734,0.0955,0.9303,0.9414,0.6797,0.9673,0.9150,0.9869,0.8085,0.9466,1,1,0.6196,0.7725,strong regression-retrieval balance
Current setting,0.2,1.0,runs/foundation/medicalnet_layer4_regalign_best.pt,0.0700,0.0894,0.9198,0.9350,0.7778,0.9804,0.9216,0.9869,0.8568,0.9493,1,1,0.6366,0.7660,current best balanced checkpoint from prior run
Strong contrastive,0.5,1.0,runs/foundation/remap_pet_layer4_cw05_best.pt,0.0766,0.0979,0.9140,0.9338,0.8366,0.9935,0.9150,0.9935,0.9015,0.9498,1,1,0.6052,0.7529,best PET-to-SUVR retrieval but weaker numeric regression
Contrastive-heavy,1.0,1.0,runs/foundation/remap_pet_layer4_cw10_best.pt,0.0935,0.1194,0.8886,0.9100,0.8170,0.9804,0.9281,0.9935,0.8886,0.9571,1,1,0.5712,0.7595,strong retrieval but clear regression degradation