Seems to be an issue with keras.ops.image.map_coordinates and TF backend. Seems to happen on both CPU and GPU.

Versions
TF: '2.19.0' Keras: '3.10.0' JAX: '0.5.2' PyTorch: '2.6.0+cu124' 
import os

os.environ["KERAS_BACKEND"] = "tensorflow"  # works for "jax" or "torch"

import keras
import matplotlib.pyplot as plt
import numpy as np

# Flag to control out-of-bounds coordinates
# Breaks TF if set to True
use_out_of_bounds = True

n_rho, n_theta = 64, 64
polar_img = np.zeros((n_rho, n_theta), dtype=np.float32)

for i in range(n_rho):
    if 8 < i < 16 or 24 < i < 32 or 40 < i < 48:
        polar_img[i, :] = 1.0

if use_out_of_bounds:
    # intentionally out-of-bounds, which break TF
    rho_idx = np.linspace(-5, n_rho + 5, n_rho)
    theta_idx = np.linspace(-5, n_theta + 5, n_theta)
else:
    rho_idx = np.linspace(0, n_rho - 1, n_rho)
    theta_idx = np.linspace(0, n_theta - 1, n_theta)

rho_grid, theta_grid = np.meshgrid(rho_idx, theta_idx, indexing="ij")
coords = np.stack([rho_grid, theta_grid], axis=0).reshape(2, -1).astype(np.float32)
coords = keras.ops.convert_to_tensor(coords)

out = keras.ops.image.map_coordinates(
    polar_img, coordinates=coords, order=0, fill_mode="constant", fill_value=0.0
)
out = keras.ops.convert_to_numpy(out).reshape(n_rho, n_theta)

plt.figure()
plt.subplot(1, 2, 1)
plt.imshow(polar_img, aspect="auto", cmap="gray")
plt.title("Input")
plt.subplot(1, 2, 2)
plt.imshow(out, aspect="auto", cmap="gray")
plt.title("map_coordinates Output")
plt.tight_layout()
plt.savefig("test.png")

Should produce (only works with torch and jax backend)

Image

But produces (with TF backend)

Image