app.py

__all__ = [
    "ORGAN",
    "IMAGE_SIZE",
    "MODEL_NAME",
    "THRESHOLD",
    "CODES",
    "learn",
    "title",
    "description",
    "examples",
    "interpretation",
    "demo",
    "x_getter",
    "y_getter",
    "splitter",
    "make3D",
    "predict",
    "infer",
    "remove_small_segs",
    "to_oberlay_image",
]

import numpy as np
import pandas as pd
import skimage
from fastai.vision.all import *
import segmentation_models_pytorch as smp

import gradio as gr

ORGAN = "kidney"
IMAGE_SIZE = 512
MODEL_NAME = "unetpp_b4_th60_d9414.pkl"
THRESHOLD = float(MODEL_NAME.split("_")[2][2:]) / 100.0
CODES = ["Background", "FTU"]  # FTU = functional tissue unit


def x_getter(r):
    return r["fnames"]


def y_getter(r):
    rle = r["rle"]
    shape = (int(r["img_height"]), int(r["img_width"]))
    return rle_decode(rle, shape).T


def splitter(model):
    enc_params = L(model.encoder.parameters())
    dec_params = L(model.decoder.parameters())
    sg_params = L(model.segmentation_head.parameters())
    untrained_params = L([*dec_params, *sg_params])
    return L([enc_params, untrained_params])


learn = load_learner(MODEL_NAME)


def make3D(t: np.array) -> np.array:
    t = np.expand_dims(t, axis=2)
    t = np.concatenate((t, t, t), axis=2)
    return t


def predict(fn, cutoff_area=200):
    data = infer(fn)
    data = remove_small_segs(data, cutoff_area=cutoff_area)
    return to_oberlay_image(data), data["df"]


def infer(fn):
    img = PILImage.create(fn)
    tf_img, _, _, preds = learn.predict(img, with_input=True)
    mask = (F.softmax(preds.float(), dim=0) > THRESHOLD).int()[1]
    mask = np.array(mask, dtype=np.uint8)
    resized_image = Image.fromarray(
        tf_img.numpy().transpose(1, 2, 0).astype(np.uint8)
    ).resize(img.shape)
    resized_image = np.array(resized_image)
    return {
        "tf_image": tf_img.numpy().transpose(1, 2, 0).astype(np.uint8),
        "tf_mask": mask,
    }


def remove_small_segs(data, cutoff_area=250):
    labeled_mask = skimage.measure.label(data["tf_mask"])
    props = skimage.measure.regionprops(labeled_mask)
    df = {"Glomerulus": [], "Area (in px)": []}
    for i, prop in enumerate(props):
        if prop.area < cutoff_area:
            labeled_mask[labeled_mask == i + 1] = 0
            continue
        df["Glomerulus"].append(len(df["Glomerulus"]) + 1)
        df["Area (in px)"].append(prop.area)
    labeled_mask[labeled_mask > 0] = 1
    data["tf_mask"] = labeled_mask.astype(np.uint8)
    data["df"] = pd.DataFrame(df)
    return data


def to_oberlay_image(data):
    img, msk = data["tf_image"], data["tf_mask"]
    msk_im = np.zeros_like(img)
    # rgb code: 255, 80, 80
    msk_im[:, :, 0] = 255
    msk_im[:, :, 1] = 80
    msk_im[:, :, 2] = 80
    img = Image.fromarray(img).convert("RGBA")
    msk_im = Image.fromarray(msk_im).convert("RGBA")
    msk = Image.fromarray((msk * 255 * 0.5).astype(np.uint8))

    img.paste(
        msk_im,
        (0, 0),
        msk,
    )
    return img


title = "Glomerulus Segmentation"
description = """
A web app that segments glomeruli in histological kidney slices!

The model deployed here is a [UNet++](https://arxiv.org/abs/1807.10165) with an [efficientnet-b4](https://arxiv.org/abs/1905.11946) encoder from the [segmentation_models_pytorch](https://github.com/qubvel/segmentation_models.pytorch) library.

The provided example images are random subset of kidney slices from the [Human Protein Atlas](https://www.proteinatlas.org/). These have been collected separately from model training and have neither been part of the training, validation nor test set.

Here is my corresponding [blog post](https://fhatje.github.io/posts/glomseg/train_model.html).
"""

examples = [str(p) for p in get_image_files("example_images")]
interpretation = "default"

demo = gr.Interface(
    fn=predict,
    inputs=gr.components.Image(width=IMAGE_SIZE, height=IMAGE_SIZE),
    outputs=[gr.components.Image(), gr.components.DataFrame()],
    title=title,
    description=description,
    examples=examples,
)

demo.launch()