Skip to content

Latest commit

 

History

History
570 lines (330 loc) · 20.2 KB

ChangeLog.md

File metadata and controls

570 lines (330 loc) · 20.2 KB
20191219

  • 图片清晰模型 dr0_0_Clear_formal.py 3rd_edition, 发现有些侧视DR图像被误判为合格, 原因是训练集中侧视图不够多, 数据增强这部分图像重新训了第三版, 并且训练的时候加入了weight和weight_decay

  • pth: /data/steven/project/Object_Detection_coastal/Classfication/output/1_img_qualification/T3_qualification_MNetV3_train_stage2_inchannel3_3edi/models/T3_qualification_MNetV3_train_stage2_inchannel3_3edi_mobilenet_v3_E39_F1_0.873612.pth

  • 结果: 线上原来错误分类为合格的侧视图都判定为不合格了

1212

  • 肺结节 更新pth和config文件, 使用drL_drZ_2019_FJJ_withBIG_clear_0.5_neg数据训练

  • /data/steven/project/Object_Detection_coastal/mmdetection_project/output/2_FJJ/FJJ_dataset_drL_drZ_2019_FJJ_withBIG_clear_0.5_neg/cascade_rcnn_dconv_c3_c5_r50_fpn_1x_alub/epoch_24.pth

  • /data/steven/project/Object_Detection_coastal/mmdetection_project/saved_cfgs/2_FJJ/drL_drZ_2019_FJJ_withBIG_clear_0.5_neg/cascade_rcnn_dconv_c3_c5_r50_fpn_1x_alub.py

  • 数据分布: drL_drZ_2019_FJJ_withBIG_clear_0.5_neg

​ name_EN t_bx v_bx a_bx t_im v_im a_im

​ FJJ 5096 224 5320 4369 190 4559

​ WYY 10494 2746 13240 10494 2746 13240

  • result: drL_drZ_0813_combine_3200_Test

FJJ_Recall FJJ_FalPos FJJ_FPsPI FJJ_Precis FJJ_img_count FJJ_pred_all FJJ_pred_r FJJ_pred_w FJJ_gt_all FJJ_gt_shot

0.863 0.825 0.943 0.175 560.0 640.0 112.0 528.0 131.0 113.0

1128

  • 肺结节 dr2_11_FeiJieJie_formal.py 加入马乐/王林的肺结节数据, 数据量更大, 在验证集和测试集上的假阳率更低, 但是在肺肿瘤测试集上的recall下降了, 后续需要进一步优化肺肿瘤问题

  • /data/steven/project/Object_Detection_coastal/mmdetection_project/output/2_FJJ/FJJ_dataset_drL_drZ_1120_6500_FJJ_mix/cascade_rcnn_dconv_c3_c5_r50_fpn_1x_alub/epoch_13.pth

  • /data/steven/project/Object_Detection_coastal/mmdetection_project/saved_cfgs/2_FJJ/drL_drZ_1120_6500_FJJ_mix/cascade_rcnn_dconv_c3_c5_r50_fpn_1x_alub.py

  • 数据分布: drL_drZ_1120_6500_FJJ_mix

name_EN t_bx v_bx a_bx t_im v_im a_im

WYY 18466 1838 20304 18466 1838 20304

FJJ 7968 319 8287 6865 288 7153

  • result:
    • drL_drZ_0813_combine_3200_Test FJJ_Recall FJJ_FalPos FJJ_FPsPI FJJ_Precis FJJ_img_count FJJ_pred_all FJJ_pred_r FJJ_pred_w FJJ_gt_all FJJ_gt_shot 0.824 0.752 0.573 0.248 560.0 427.0 106.0 321.0 131.0 108.0
    • drL_drZ_1120_6500_FJJ_mix_val FJJ_Recall FJJ_FalPos FJJ_FPsPI FJJ_Precis FJJ_img_count FJJ_pred_all FJJ_pred_r FJJ_pred_w FJJ_gt_all FJJ_gt_shot 0.727 0.809 0.538 0.191 1838.0 1222.0 234.0 988.0 319.0 232.0
1102

  • 4合一 dr1_11_SC_FQZ_FBZ_SZGBK_formal.py 发现将4in1拆开成2个模型跟直接合并预测的结果差不多, 只要4in1进行了合适的权重设置(占比倒数)

  • /data/steven/project/Object_Detection_coastal/Classfication/output/9_4in1/T10_4in1_MNetV3_train_stage2_inchannel3_BCELoss_best_weights/models/T10_4in1_MNetV3_train_stage2_inchannel3_BCELoss_best_weights_mobilenet_v3_E32.pth

  • 数据分布: T10_bestF1_best_weights

9_4in1 train val

渗出 1307 158

肺气肿 545 123

肺不张 108 30

上纵膈变宽 1755 225

纯阴性 1802 190

图片总数 5390 700

  • result: T10_bestF1_best_weights

SC FQZ FBZ SZGBK mean_f1

recall fp recall fp recall fp recall fp

0.80 0.33 0.75 0.22 0.46 0.36 0.74 0.21 0.70

1031

  • dr1_11_FQZ_FBZ_formal.py|dr1_12_SC_SZGBK_formal.py 将FQZ、FBZ|SC、SZGBK分别用两个二合一模型预测, 因为这两对的训练数据量比较接近, 没有数据不平衡的问题.

  • 尝试过3个或4个病种合并预测, 数据不平衡的问题对结果会有影响. 尝试过复制、random_crop、上采样、下采样等数据平衡方法, 效果不明显. 最有效的还是把数据量差异大的病种分开预测.

  • 其中FQZ_FBZ模型尝试过混入NIHCC公开数据集进行训练, 发现公开数据集的数据一致性很差, 混入训练会影响模型的结果, 放弃

  • /data/steven/project/Object_Detection_coastal/Classfication/output/11_FQZ_FBZ/T2.4_FQZ_FBZ_MNetV3_train_without_outerdata_added_more_neg_val/models/T2.4_FQZ_FBZ_MNetV3_train_without_outerdata_added_more_neg_val_mobilenet_v3_E40.pth

  • /data/steven/project/Object_Detection_coastal/Classfication/output/12_SC_SZGBK/T1_SC_SZGBK_MNetV3_train_without_outerdata_added/models/T1_SC_SZGBK_MNetV3_train_without_outerdata_added_mobilenet_v3_E14.pth

  • 数据分布:

T2.4_without_outerdata_added_more_neg_val

11_FQZ_FBZ train val

单肺气肿 542 120

单肺不张 106 28

肺气肿+肺不张 1 2

纯阴性 1000 550

总数 1649 700

T1_without_outerdata_added 12_SC_SZGBK train val 单渗出 1225 148 单上纵膈变宽 1672 215 渗出+上纵膈变宽 81 10 纯阴性 2000 327 总数 4978 700

  • result: T2.4_without_outerdata_added_more_neg_val

FQZ FBZ mean_f1

recall fp recall fp

0.78 0.29 0.43 0.48 0.61

  • result: T2.4_without_outerdata_added_more_neg_val

SC SZGBK mean_f1

recall fp recall fp

0.79 0.37 0.78 0.26 0.73

20191028
  • dr0_0_Clear_formal.py 上一版将一小部分合格的图片判定成不合格, 原因是训练用的不合格数据包含了合格数据(不合格定义不严), 重新筛选不合格数据, 然后重新训练了分类模型
  • /data/steven/project/Object_Detection_coastal/Classfication/output/1_img_qualification/T2_qualification_MNetV3_train_stage2_inchannel3_2edi/models/T2_qualification_MNetV3_train_stage2_inchannel3_2edi_mobilenet_v3_E25_F1_0.855359.pth
20191024

  • dr0_0_Clear_formal.py 重新上线了一个分类模型, 用于图片是否合格的判定, 0:不确定, 1:合格, 2:不合格

  • /data/steven/project/Object_Detection_coastal/Classfication/output/1_img_qualification/T1_qualification_MNetV3_train_stage2_inchannel3/models/T1_qualification_MNetV3_train_stage2_inchannel3_mobilenet_v3_E39_F1_0.858925.pth

  • 数据分布:

1_img_qualification Train val notsure 818 91 qualified 5702 634 unqualified 1242 139 all 7762 864

  • result:

​ notsure qualified unqualified mean_f1 ​ recall fp recall fp recall fp T1_benchmark 0.57 0.22 0.98 0.05 0.95 0.05 0.86

20191021
  • dr2_10_RuTouYing_formal.py 只有单侧预测出乳头影,将其修改为肺结节,模型不变
20191021

  • 肺气肿、肺不张、上纵膈变宽, 3合1, dr1_11_SC_FQZ_FBZ_SZGBK_formal.py 去掉渗出, 训练数据减少了接近一半, 数据稍微有点不平衡, 效果相比4合1得到提升

  • /data/steven/project/Object_Detection_coastal/Classfication/output/10_3in1/T1_3in1_MNetV3_train_stage2_inchannel3_valthresh_bestf1/models/T1_3in1_MNetV3_train_stage2_inchannel3_valthresh_bestf1_mobilenet_v3_E37.pth

  • 数据分布:

9_3in1 Train val 肺气肿 544 121 肺不张 107 30 上纵膈变宽 1773 205 纯阴性 1792 200 all 4216 556

  • result:

​ FQZ FBZ SZGBK mean_f1 ​ recall fp recall fp recall fp T1_benchmark 0.75 0.16 0.60 0.35 0.90 0.16 0.76

20191018
  • 渗出、肺气肿、肺不张、上纵膈变宽, 4合1,dr1_11_SC_FQZ_FBZ_SZGBK_formal.py, 肺气肿和肺不张数据量少, 数据严重不平衡, 效果不够好
  • /data/steven/project/Object_Detection_coastal/Classfication/output/9_4in1/T5_4in1_MNetV3_train_stage2_inchannel3_valthresh_bestf1/models/T5_4in1_MNetV3_train_stage2_inchannel3_valthresh_bestf1_mobilenet_v3_E29.pth
20191015
  • 胸膜增厚第二版

ID Name_CN Name_EN box_Count img_Count 1 无意义 WYY 3170 3170 2 胸膜增厚 XMZH 5502 2463 3 胸腔积液 XQJY 969 764

  • config
    • /data/steven/project/Object_Detection_coastal/mmdetection_project/saved_cfgs/5_XMZH_JY/XMZH_JY_1010_3170_crop_stage2_expand_upd_T2/cascade_rcnn_dconv_c3_c5_r50_fpn_1x_num_libra_alub.py
  • pth
    • /data/steven/project/Object_Detection_coastal/mmdetection_project/output/5_XMZH_JY/XMZH_JY_dataset_XMZH_JY_1010_3170_crop_stage2_expand_upd_T2/cascade_rcnn_dconv_c3_c5_r50_fpn_1x_num_libra_alub/epoch_29.pth
  • 结果
    • ​ XMZH JY ​ val test gt516 val test gt96 ​ epoch FalPos r FalPos r 漏 多 epoch FalPos r FalPos r 漏 多 XMZH_JY_0924_1222_crop_stage2_T2 29 0.60 0.90 0.33 0.92 44 719 34 0.62 0.90 0.37 0.83 16 48
20191014
  • 器官 器官和椎体合并 959张标注
  • config: /data/steven/project/Object_Detection_coastal/mmdetection_project/saved_cfgs/4_QiGuan_Series/1011_ZuiTi_QiGuan_960_T0/cascade_mask_rcnn_dconv_c3-c5_r50_fpn_1x.py
  • pth /data/steven/project/Object_Detection_coastal/mmdetection_project/output/10_QiGuan_series/ZuiTi_QiGuan_dataset_1011_ZuiTi_QiGuan_960_T0/cascade_mask_rcnn_dconv_c3-c5_r50_fpn_1x/epoch_34.pth
20191010
  • 乳头影 dr2_10_RuTouYing_formal 增加了客户端分布的标注图片, 提高了模型的泛化性; 修正了一个cut_thresh=0.3下的bug, in_norner_outer_bbox()不再需要outer_bbox进行过滤
  • /data/steven/project/Object_Detection_coastal/mmdetection_project/output/0_RTY/RTY_dataset_RTY_1009_4000_combine_Stage2_up020_4000_T3/cascade_rcnn_dconv_c3_c5_r50_fpn_1x_num_libra_alub/epoch_34.pth
  • /data/steven/project/Object_Detection_coastal/mmdetection_project/saved_cfgs/0_RTY/RTY_1009_4000_combine_Stage2_up020_4000_T3/cascade_rcnn_dconv_c3_c5_r50_fpn_1x_num_libra_alub.py
20190927

  • 胸膜增厚,积液 dr2_13_XMZH_JY_formal: thresh 0.3

  • /data/steven/project/Object_Detection_coastal/mmdetection_project/output/5_XMZH_JY/XMZH_JY_dataset_XMZH_JY_0924_1222_crop_stage2_T0/cascade_rcnn_dconv_c3_c5_r50_fpn_1x_num_libra_alub/epoch_33.pth

  • XMZH_JY_0924_1222_crop_stage2_T1 epoch 33

name_EN t_bx v_bx a_bx t_im v_im a_im

XMZH 2636 194 2830 2096 148 2244

XQJY 480 40 520 480 40 520

WYY 3730 210 3940 3730 210 3940

  • 胸膜增厚

val 0.56 0.90 test 0.53 0.83 (0.3)

  • 积液

val 0.28 0.90 test 0.33 0.803 83 (0.3)

20190925
  • 肋骨 dr1_1_LeiGu_formal 用mobilenetv1训练了一个找肋骨的目标检测模型代替原有的重量级mask_rcnn实例分割模型, 因为找肋骨的目的是为了确定第2、第4根肋骨的下沿线进行肺的上中下野定位, 轻量级的目标检测即可完成, 显存下降7-8G
  • /data/chenwh/mobilenet_faster_rcnn/output/mobile/11_leigu/LG_6cls_clahe40_mixed_train/default/mobile_faster_rcnn_iter_80000.ckpt

09_23_6cls_clahe40_mixed Name_CN name_EN t_bx v_bx a_bx t_im v_im a_im 左肋骨_1 ZLG_1 1683 297 1980 1683 297 1980 左肋骨_2 ZLG_2 1683 297 1980 1683 297 1980 左肋骨_3 ZLG_3 1683 297 1980 1683 297 1980 左肋骨_4 ZLG_4 1683 297 1980 1683 297 1980 左肋骨_5 ZLG_5 1683 297 1980 1683 297 1980 左肋骨_6 ZLG_6 1683 297 1980 1683 297 1980 右肋骨_1 YLG_1 1683 297 1980 1683 297 1980 右肋骨_2 YLG_2 1683 297 1980 1683 297 1980 右肋骨_3 YLG_3 1683 297 1980 1683 297 1980 右肋骨_4 YLG_4 1683 297 1980 1683 297 1980 右肋骨_5 YLG_5 1683 297 1980 1683 297 1980 右肋骨_6 YLG_6 1683 297 1980 1683 297 1980

  • 肺结节 dr2_11_FeiJieJie_formal 添加肺肿瘤标注(大号肺结节), 原有的肺结节标注及参数不变, 重新训练了模型, 目的是解决漏诊肺肿瘤的情况. 目前出现了一些肺肿瘤假阳, 待解决

  • /data/steven/project/Object_Detection_coastal/mmdetection_project/output/2_FJJ/FJJ_dataset_drL_drZ_0921_5100_FJJ_FZK_T0/cascade_rcnn_dconv_c3_c5_r50_fpn_1x_alub/epoch_11.pth

​ val 0.86 0.78 ​ test 0.81 0.82 ​ testset: ​ /data/steven/project/Object_Detection_coastal/dataser_raw/3_feijiejie/COCOjson/train_val/drL_drZ_0813_combine_3200_stage2_T0_test.json ​ thresh:0.1, gt:131, miss:23, extra:477 ​ thresh:0.2, gt:131, miss:31, extra:292

  • drL_drZ_0921_5100_FJJ_FZK_T0

name_EN t_bx v_bx a_bx t_im v_im a_im WYY 11161 1374 12535 11161 1374 12535 FJJ 5077 210 5287 4349 189 4538

20190918
  • 乳头影 dr2_10_RuTouYing_formal rutouying_crop框太上了, 导致较多误诊, 修改8in1模型up0.2,down0(原up0.52,down0.13), 并且拿cut_thresh_w=0.3裁剪之后的图片去训练(原训练图片没有经过cut_thresh裁剪)
  • /data/steven/project/Object_Detection_coastal/mmdetection_project/output/0_RTY/RTY_dataset_RTY_0829_2400_AIcombine_Stage2_up020_T8/cascade_rcnn_dconv_c3_c5_r50_fpn_1x_num_libra_alub/epoch_30.pth

val 0.19 0.93

test 0.23 0.92

testset:

​ /data/steven/project/Object_Detection_coastal/dataser_raw/1_rutouying/COCOjson/train_val/RTY_0829_2400_AIcombine_T6_test.json

​ thresh 0.1

​ gt:151 miss:11 extra:42

  • RTY_0829_2400_AIcombine_Stage2_up020_T8

name_EN t_bx v_bx a_bx t_im v_im a_im RTY 2967 360 3327 2967 360 3327 WYY 5752 842 6594 5752 842 6594

  • 斑片影 dr2_3_BanPianYing_formal 增加alub提高泛化能力, 增加数据量, 总数据量达到3300

  • /data/steven/project/Object_Detection_coastal/mmdetection_project/output/1_BPY/BPY_dataset_BPY_0913_3300_crop_stage2_T1/cascade_rcnn_dconv_c3_c5_r50_fpn_1x_alub/epoch_21.pth val 0.85 0.8 test 0.84 0.83 testset: /data/steven/project/Object_Detection_coastal/dataser_raw/2_banpianying/COCOjson/train_val/BPY_0903_2800_crop_stage2_T1_test.json thresh 0.1 gt:192 miss:32 extra:994

  • BPY_0913_3300_crop_stage2_T1

name_EN t_bx v_bx a_bx t_im v_im a_im WYY 6604 308 6912 6604 308 6912 BPY 4627 250 4877 3334 186 3520

20190911
  • 血管壁钙化
    • dr2_5_XueGuanBiGaiHua_formal
    • /data/steven/project/Object_Detection_coastal/mmdetection_project/output/3_XGBGH/XGB_dataset_XGBGH_0820_2900_2stage_T0a_Alltrain/cascade_rcnn_dconv_c3_c5_r50_fpn_1x_num_libra_alub/epoch_22.pth

​ val 0.28 0.89

​ test 0.37 0.90

​ XGBGH_0820_2900_2stage_T0a_Alltrain

​ name_EN t_bx v_bx a_bx t_im v_im a_im

​ XGBGH 4095 343 4438 3233 270 3503

​ WYY 6756 564 7320 6756 564 7320

  • 乳头影
    • dr2_10_RuTouYing_formal 训练的时候增加alub处理, 随机改变明暗度, 提高模型泛化能力, 增加了迭代次数, 调整了学习率衰减策略
    • /data/steven/project/Object_Detection_coastal/mmdetection_project/output/0_RTY/RTY_dataset_RTY_0829_2400_AIcombine_Stage2_T6/cascade_rcnn_dconv_c3_c5_r50_fpn_1x_num_libra_alub_2x/epoch_24.pth

​ val 0.18 0.88

​ test 0.16 0.88

​ testset:

​ /data/steven/project/Object_Detection_coastal/dataser_raw/1_rutouying/COCOjson/train_val/RTY_0829_2400_AIcombine_T6_test.json

​ thresh 0.1

​ with corner_outer_filter

​ gt:151 miss:18 extra:25

RTY_0829_2400_AIcombine_Stage2_T6 expand						
name_EN   	t_bx 	v_bx 	a_bx 	t_im 	v_im 	a_im
WYY       	8404	842		9246	8404	842		9246
RTY       	4527	376		4903	4527	376		4903
  • 肺结节
    • dr2_11_FeiJieJie_formal
    • /data/steven/project/Object_Detection_coastal/mmdetection_project/output/2_FJJ/FJJ_dataset_drL_drZ_0905_4200_allAI_pred_T3/cascade_rcnn_dconv_c3_c5_r50_fpn_1x_alub/epoch_16.pth

​ val 0.90 0.73

​ test 0.82 0.79

​ testset:

​ /data/steven/project/Object_Detection_coastal/dataser_raw/3_feijiejie/COCOjson/train_val/drL_drZ_0813_combine_3200_stage2_T0_test.json

​ thresh 0.1 gt:131 miss:27 extra:473

​ thresh 0.2 gt:131 miss:31 extra:311

drL_drZ_0905_4200_allAI_pred_T3						
name_EN   	t_bx 	v_bx 	a_bx 	t_im 	v_im 	a_im
FJJ       	3431	148		3579	2967	135		3102
WYY       	8520	1256	9776	8520	1256	9776
  • 斑片影
    • dr2_3_BanPianYing_formal 增加alub
    • /data/steven/project/Object_Detection_coastal/mmdetection_project/output/1_BPY/BPY_dataset_BPY_0903_2800_crop_stage2_T1/cascade_rcnn_dconv_c3_c5_r50_fpn_1x_alub/epoch_30.pth

​ val 0.76 0.68

​ test 0.75 0.72

​ testset:

​ /data/steven/project/Object_Detection_coastal/dataser_raw/2_banpianying/COCOjson/train_val/BPY_0903_2800_crop_stage2_T1_test.json

​ thresh 0.1 gt:192 miss:54 extra:461

BPY_0903_2800_crop_stage2_T1						
name_EN   	t_bx 	v_bx 	a_bx 	t_im 	v_im 	a_im
WYY       	1866	308		2174	1866	308		2174
BPY       	1554	204		1758	1272	159		1431
  • 肋软骨钙化
    • dr2_12_LeiRuanGuGaiHua_formal
    • /data/chenwh/mobilenet_faster_rcnn/output/mobile/10_leiruangugaihua/0910_500_LRGgaihua/default/mobile_faster_rcnn_iter_30000.ckpt

​ val 0.25 0.982

​ testset:

​ /data/steven/project/Object_Detection_coastal/dataser_raw/11_LRG_gaihua/COCOjson/train_val/0910_500_LRGgaihua_val.json

​ thresh 0.1 gt:192 miss:3 extra:58

0910_500_LRGgaihua_T0						
name_EN   	t_bx 	v_bx 	a_bx 	t_im 	v_im 	a_im 
WYY       	1818	232	2050	1818	232	2050
LRGGH     	1303	171	1474	1299	171	1470
20190905
  • 斑片影
    • dr2_3_BanPianYing_formal
    • /data/steven/project/Object_Detection_coastal/mmdetection_project/output/1_BPY/BPY_dataset_BPY_0903_2800_crop_stage2_T1/cascade_rcnn_dconv_c3_c5_r50_fpn_1x_num_libra/epoch_18.pth

​ val 0.75 0.64

​ test 0.73 0.64

​ testset:

​ '/data/steven/project/Object_Detection_coastal/dataser_raw/2_banpianying/COCOjson/train_val/BPY_0903_2800_crop_stage2_T1_test.json'

​ thresh 0.1

​ gt:192 miss:69 extra:347

BPY_0903_2800_crop_stage2_T1						
name_EN   	t_bx 	v_bx 	a_bx 	t_im 	v_im 	a_im
WYY       	1866	308		2174	1866	308		2174
BPY       	1554	204		1758	1272	159		1431
  • 乳头影
    • dr2_10_RuTouYing_formal
    • /data/steven/project/Object_Detection_coastal/mmdetection_project/output/0_RTY/RTY_dataset_RTY_0829_2400_AIcombine_Stage2_T6/cascade_rcnn_dconv_c3_c5_r50_fpn_1x_num_libra/epoch_23.pth

​ val 0.23 0.91

​ test 0.31 0.91

​ testset:

​ /data/steven/project/Object_Detection_coastal/dataser_raw/1_rutouying/COCOjson/train_val/RTY_0829_2400_AIcombine_T6_test.json

​ thresh 0.2

​ with corner_outer_filter

​ gt:151 miss:14 extra:63

RTY_0829_2400_AIcombine_Stage2_T6 expand						
name_EN   	t_bx 	v_bx 	a_bx 	t_im 	v_im 	a_im
WYY       	8404	842		9246	8404	842		9246
RTY       	4527	376		4903	4527	376		4903
20190822

  • dr1_10_8in1_crop_formal.ckpt

    • 8合1模型, 共用了650张图片
    • 相比前一版增加了50多张新图片及标注进行训练,
    • 微调了左右肺, 左右肺门及乳头影的bbox框位置, 使更加准确
    • 上一版有些标注肺门在同一侧, 或者bbox缺失, 这次都补上了

  • 肺结节

    • dr2_11_FeiJieJie_formal
    • /data/steven/project/Object_Detection_coastal/mmdetection_project/output/2_FJJ/FJJ_drLin_1068_7_22_drL_drZ_0813_combine_3200_stage2_random_erasing_T1/cascade_rcnn_dconv_c3_c5_r50_fpn_1x/epoch_5.pth
    • 使用 drL_drZ_0813_combine_3200_stage2_random_erasing_T1 epoch5

​ val 0.845 0.733

​ test 0.884 0.824

​ thre 0.1 input large img_scale=(1900, 1200),

name_EN   	t_bx 	v_bx 	a_bx 	t_im 	v_im 	a_im
	WYY       	9288	776	10064	9288	776	10064
	FJJ       	3860	240	4100	3340	211	3551
  • 班片影
    • dr2_3_BanPianYing_formal
    • /data/steven/project/Object_Detection_coastal/mmdetection_project/output/1_BPY/BPY_DRl_1799_7_31_1946_0817_stage2_crop_T1/cascade_rcnn_dconv_c3_c5_r50_fpn_1x/epoch_5.pth
    • 使用 BPY_1946_0817_crop_stage2_comb_T1 cascade_rcnn_dconv_c3_c5_r50_fpn_1x_exg epoch5

​ val 0.91 0.81

​ test 0.906 0.791

name_EN   	t_bx 	v_bx 	a_bx 	t_im 	v_im 	a_im
BPY       	2628	105	2733	2145	80	2225
WYY       	2516	452	2968	2516	452	2968
  • 乳头影
    • dr2_10_RuTouYing_formal
    • /data/steven/project/Object_Detection_coastal/mmdetection_project/output/0_RTY/RTY_drLin_805_7_22_stage2_with_clahe_0803_2000_half_8in1_T0/cascade_rcnn_dconv_c3-c5_r50_fpn_1x/epoch_24.pth
    • 使用 RTY_drLin_805_7_22_stage2_with_clahe_0803_2000_half_8in1_T0 epoch24

​ val 0.20 0.91

​ test 0.205 0.848

name_EN   	t_bx 	v_bx 	a_bx 	t_im 	v_im 	a_im
WYY       	11160	700	11860	11160	700	11860
RTY       	4103	278	4381	4103	278	4381
  • 血管壁钙化
    • dr2_5_XueGuanBiGaiHua_formal
    • /data/steven/project/Object_Detection_coastal/mmdetection_project/output/3_XGBGH/XGB_drLin_904_Stage2_XGBGH_0820_2900_2stage_T0c_notAlltrain/cascade_rcnn_dconv_c3_c5_r50_fpn_1x/epoch_7.pth'
    • 使用XGBGH_0820_2900_2stage_T0c_notAlltrain _new8in1 epoch7

​ val 0.39 0.92

​ test 0.521 0.922

name_EN   	t_bx 	v_bx 	a_bx 	t_im 	v_im 	a_im
	XGBGH     	2266	376	2642	1752	294	2046
	WYY       	3681	616	4297	3681	616	4297