Commit
•
68542c0
1
Parent(s):
73878cb
Update README.md (#4)
Browse files- Update README.md (0bd301187780ab424b9c9b3e321382d464447127)
Co-authored-by: Ziyue Wang <ZiyueWang@users.noreply.huggingface.co>
README.md
CHANGED
@@ -19,6 +19,58 @@ OpticDisc (Fundus Image)
|
|
19 |
|
20 |
Thyroid Nodule (UltraSound)
|
21 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
22 |
Download the Adapters you need [here](https://huggingface.co/KidsWithTokens/Medical-Adapter-Zoo/tree/main)
|
23 |
|
24 |
## What
|
@@ -29,7 +81,7 @@ Check our paper: [Medical SAM Adapter](https://arxiv.org/abs/2304.12620) for the
|
|
29 |
|
30 |
SAM (Segment Anything Model) is one of the most popular open models for image segmentation. Unfortunately, it does not perform well on the medical images.
|
31 |
An efficient way to solve it is using Adapters, i.e., some layers with a few parameters to be added to the pre-trained SAM model to fine-tune it to the target down-stream tasks.
|
32 |
-
Medical image segmentation includes many different organs, lesions, abnormalities as the targets.
|
33 |
So we are training different adapters for each of the targets, and sharing them here for the easy usage in the community.
|
34 |
|
35 |
Download an adapter for your target disease—trained on organs, lesions, and abnormalities—and effortlessly enhance SAM.
|
@@ -55,14 +107,34 @@ GPUdevice = torch.device('cuda', args.gpu_device)
|
|
55 |
|
56 |
# load the original SAM model
|
57 |
net = get_network(args, args.net, use_gpu=args.gpu, gpu_device=GPUdevice, distribution = args.distributed)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
58 |
|
59 |
# load task-specific adapter
|
60 |
adapter_path = 'OpticCup_Fundus_SAM_1024.pth'
|
61 |
-
|
62 |
-
|
63 |
-
|
64 |
-
|
65 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
66 |
|
67 |
## Authorship
|
68 |
|
|
|
19 |
|
20 |
Thyroid Nodule (UltraSound)
|
21 |
|
22 |
+
Aorta (Abdominal Image)
|
23 |
+
|
24 |
+
Esophagus (Abdominal Image)
|
25 |
+
|
26 |
+
Gallbladder (Abdominal Image)
|
27 |
+
|
28 |
+
Inferior Vena Cava (Abdominal Image)
|
29 |
+
|
30 |
+
Left Adrenal Gland (Abdominal Image)
|
31 |
+
|
32 |
+
Right Adrenal Gland (Abdominal Image)
|
33 |
+
|
34 |
+
Left Kidney (Abdominal Image)
|
35 |
+
|
36 |
+
Right kidney (Abdominal Image)
|
37 |
+
|
38 |
+
Liver (Abdominal Image)
|
39 |
+
|
40 |
+
Pancreas (Abdominal Image)
|
41 |
+
|
42 |
+
Spleen (Abdominal Image)
|
43 |
+
|
44 |
+
Stomach (Abdominal Image)
|
45 |
+
|
46 |
+
Portal Vein and Splenic Vein (Abdominal Image)
|
47 |
+
|
48 |
+
Edematous Tissue (Brain Tumor mpMRI)
|
49 |
+
|
50 |
+
Enhancing Tumor (Brain Tumor mpMRI)
|
51 |
+
|
52 |
+
Necrotic (Brain Tumor mpMRI)
|
53 |
+
|
54 |
+
Inferior Alveolar Nerve (CBCT)
|
55 |
+
|
56 |
+
Instrument Clasper (Surgical Video)
|
57 |
+
|
58 |
+
Instrument Shaft (Surgical Video)
|
59 |
+
|
60 |
+
Instrument Wrist (Surgical Video)
|
61 |
+
|
62 |
+
Kidney Tumor (MRI)
|
63 |
+
|
64 |
+
Liver (Liver Tumor CE-MRI)
|
65 |
+
|
66 |
+
Tumor (Liver Tumor CE-MRI)
|
67 |
+
|
68 |
+
Mandible (XRay)
|
69 |
+
|
70 |
+
Retina Vessel (Fundus Image)
|
71 |
+
|
72 |
+
White Blood Cell (MicroScope)
|
73 |
+
|
74 |
Download the Adapters you need [here](https://huggingface.co/KidsWithTokens/Medical-Adapter-Zoo/tree/main)
|
75 |
|
76 |
## What
|
|
|
81 |
|
82 |
SAM (Segment Anything Model) is one of the most popular open models for image segmentation. Unfortunately, it does not perform well on the medical images.
|
83 |
An efficient way to solve it is using Adapters, i.e., some layers with a few parameters to be added to the pre-trained SAM model to fine-tune it to the target down-stream tasks.
|
84 |
+
Medical image segmentation includes many different organs, lesions, and abnormalities as the targets.
|
85 |
So we are training different adapters for each of the targets, and sharing them here for the easy usage in the community.
|
86 |
|
87 |
Download an adapter for your target disease—trained on organs, lesions, and abnormalities—and effortlessly enhance SAM.
|
|
|
107 |
|
108 |
# load the original SAM model
|
109 |
net = get_network(args, args.net, use_gpu=args.gpu, gpu_device=GPUdevice, distribution = args.distributed)
|
110 |
+
net.eval()
|
111 |
+
|
112 |
+
sam_weights = 'checkpoint/sam/sam_vit_b_01ec64.pth' # load the original SAM weight
|
113 |
+
with open(sam_weights, "rb") as f:
|
114 |
+
state_dict = torch.load(f)
|
115 |
+
new_state_dict = {k: v for k, v in state_dict.items() if k in net.state_dict() and net.state_dict()[k].shape == v.shape}
|
116 |
+
net.load_state_dict(new_state_dict, strict = False)
|
117 |
|
118 |
# load task-specific adapter
|
119 |
adapter_path = 'OpticCup_Fundus_SAM_1024.pth'
|
120 |
+
checkpoint_file = os.path.join(adapter_path)
|
121 |
+
assert os.path.exists(checkpoint_file)
|
122 |
+
loc = 'cuda:{}'.format(args.gpu_device)
|
123 |
+
checkpoint = torch.load(checkpoint_file, map_location=loc)
|
124 |
+
|
125 |
+
state_dict = checkpoint['state_dict']
|
126 |
+
if args.distributed != 'none':
|
127 |
+
from collections import OrderedDict
|
128 |
+
new_state_dict = OrderedDict()
|
129 |
+
for k, v in state_dict.items():
|
130 |
+
# name = k[7:] # remove `module.`
|
131 |
+
name = 'module.' + k
|
132 |
+
new_state_dict[name] = v
|
133 |
+
# load params
|
134 |
+
else:
|
135 |
+
new_state_dict = state_dict
|
136 |
+
|
137 |
+
net.load_state_dict(new_state_dict,strict = False)
|
138 |
|
139 |
## Authorship
|
140 |
|